RL-practice-Expected Sarsa.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "2926c390f05ffdd6b94b3d79b5c0b494",
     "grade": false,
     "grade_id": "cell1",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "# Assignment 2 - Implement your agent\n",
    "\n",
    "Welcome to Course 4, Programming Assignment 2! We have learned about reinforcement learning algorithms for prediction and control in previous courses and extended those algorithms to large state spaces using function approximation. One example of this was in assignment 2 of course 3 where we implemented semi-gradient TD for prediction and used a neural network as the function approximator. In this notebook, we will build a reinforcement learning agent for control, again using a neural network for function approximation. This combination of neural network function approximators and reinforcement learning algorithms, often referred to as Deep RL, is an active area of research and has led to many impressive results (e. g., AlphaGo: https://deepmind.com/research/case-studies/alphago-the-story-so-far).\n",
    "\n",
    "**In this assignment, you will:**\n",
    "  1. Extend the neural network code from assignment 2 of course 3 to output action-values instead of state-values.\n",
    "  2. Write up the Adam algorithm for neural network optimization.\n",
    "  3. Understand experience replay buffers.\n",
    "  4. Implement Softmax action-selection.\n",
    "  5. Build an Expected Sarsa agent by putting all the pieces together.\n",
    "  6. Solve Lunar Lander with your agent."
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "d144c4fd87d2731e3fc9234172decb79",
     "grade": false,
     "grade_id": "cell2",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Packages\n",
    "- [numpy](www.numpy.org) : Fundamental package for scientific computing with Python.\n",
    "- [matplotlib](http://matplotlib.org) : Library for plotting graphs in Python.\n",
    "- [RL-Glue](http://www.jmlr.org/papers/v10/tanner09a.html), BaseEnvironment, BaseAgent : Library and abstract classes to inherit from  for reinforcement learning experiments.\n",
    "- [LunarLanderEnvironment](https://gym.openai.com/envs/LunarLander-v2/) : An RLGlue environment that wraps a LundarLander environment implementation from OpenAI Gym.\n",
    "- [collections.deque](https://docs.python.org/3/library/collections.html#collections.deque): a double-ended queue implementation. We use deque to implement the experience replay buffer.\n",
    "- [copy.deepcopy](https://docs.python.org/3/library/copy.html#copy.deepcopy): As objects are not passed by value in python, we often need to make copies of mutable objects. copy.deepcopy allows us to make a new object with the same contents as another object. (Take a look at this link if you are interested to learn more: https://robertheaton.com/2014/02/09/pythons-pass-by-object-reference-as-explained-by-philip-k-dick/)\n",
    "- [tqdm](https://github.com/tqdm/tqdm) : A package to display progress bar when running experiments\n",
    "- [os](https://docs.python.org/3/library/os.html): Package used to interface with the operating system. Here we use it for creating a results folder when it does not exist.\n",
    "- [shutil](https://docs.python.org/3/library/shutil.html): Package used to operate on files and folders. Here we use it for creating a zip file of the results folder.\n",
    "- plot_script: Used for plotting learning curves using matplotlib."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "# Import necessary libraries\n",
    "# DO NOT IMPORT OTHER LIBRARIES - This will break the autograder.\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline\n",
    "\n",
    "from rl_glue import RLGlue\n",
    "from environment import BaseEnvironment\n",
    "\n",
    "from lunar_lander import LunarLanderEnvironment\n",
    "\n",
    "from agent import BaseAgent\n",
    "\n",
    "from collections import deque\n",
    "\n",
    "from copy import deepcopy\n",
    "\n",
    "from tqdm import tqdm\n",
    "\n",
    "import os \n",
    "\n",
    "import shutil\n",
    "\n",
    "from plot_script import plot_result"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "3528df9f07a2c2a27c11a47d94849931",
     "grade": false,
     "grade_id": "cell4",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 1: Action-Value Network\n",
    "This section includes the function approximator that we use in our agent, a neural network. In Course 3 Assignment 2, we used a neural network as the function approximator for a policy evaluation problem. In this assignment, we will use a neural network for approximating the action-value function in a control problem. The main difference between approximating a state-value function and an action-value function using a neural network is that in the former the output layer only includes one unit whereas in the latter the output layer includes as many units as the number of actions. \n",
    "\n",
    "In the cell below, you will specify the architecture of the action-value neural network. More specifically, you will specify `self.layer_size` in the `__init__()` function. \n",
    "\n",
    "We have already provided `get_action_values()` and `get_TD_update()` methods. The former computes the action-value function by doing a forward pass and the latter computes the gradient of the action-value function with respect to the weights times the TD error. These `get_action_values()` and `get_TD_update()` methods are similar to the `get_value()` and `get_gradient()` methods that you implemented in Course 3 Assignment 2. The main difference is that in this notebook, they are designed to be applied to batches of states instead of one state. You will later use these functions for implementing the agent."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Work Required: Yes. Fill in the code for layer_sizes in __init__ (~1 Line). \n",
    "# Also go through the rest of the code to ensure your understanding is correct.\n",
    "class ActionValueNetwork:\n",
    "    # Work Required: Yes. Fill in the layer_sizes member variable (~1 Line).\n",
    "    def __init__(self, network_config):\n",
    "        self.state_dim = network_config.get(\"state_dim\")\n",
    "        self.num_hidden_units = network_config.get(\"num_hidden_units\")\n",
    "        self.num_actions = network_config.get(\"num_actions\")\n",
    "        \n",
    "        self.rand_generator = np.random.RandomState(network_config.get(\"seed\"))\n",
    "        \n",
    "        # Specify self.layer_size which shows the number of nodes in each layer\n",
    "        ### START CODE HERE (~1 Line)\n",
    "        self.layer_sizes = (self.state_dim, self.num_hidden_units, self.num_actions)\n",
    "        ### END CODE HERE\n",
    "        \n",
    "        # Initialize the weights of the neural network\n",
    "        # self.weights is an array of dictionaries with each dictionary corresponding to \n",
    "        # the weights from one layer to the next. Each dictionary includes W and b\n",
    "        self.weights = [dict() for i in range(0, len(self.layer_sizes) - 1)]\n",
    "        for i in range(0, len(self.layer_sizes) - 1):\n",
    "            self.weights[i]['W'] = self.init_saxe(self.layer_sizes[i], self.layer_sizes[i + 1])\n",
    "            self.weights[i]['b'] = np.zeros((1, self.layer_sizes[i + 1]))\n",
    "    \n",
    "    # Work Required: No.\n",
    "    def get_action_values(self, s):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            s (Numpy array): The state.\n",
    "        Returns:\n",
    "            The action-values (Numpy array) calculated using the network's weights.\n",
    "        \"\"\"\n",
    "        \n",
    "        W0, b0 = self.weights[0]['W'], self.weights[0]['b']\n",
    "        psi = np.dot(s, W0) + b0\n",
    "        x = np.maximum(psi, 0)\n",
    "        \n",
    "        W1, b1 = self.weights[1]['W'], self.weights[1]['b']\n",
    "        q_vals = np.dot(x, W1) + b1\n",
    "\n",
    "        return q_vals\n",
    "    \n",
    "    # Work Required: No.\n",
    "    def get_TD_update(self, s, delta_mat):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            s (Numpy array): The state.\n",
    "            delta_mat (Numpy array): A 2D array of shape (batch_size, num_actions). Each row of delta_mat  \n",
    "            correspond to one state in the batch. Each row has only one non-zero element \n",
    "            which is the TD-error corresponding to the action taken.\n",
    "        Returns:\n",
    "            The TD update (Array of dictionaries with gradient times TD errors) for the network's weights\n",
    "        \"\"\"\n",
    "\n",
    "        W0, b0 = self.weights[0]['W'], self.weights[0]['b']\n",
    "        W1, b1 = self.weights[1]['W'], self.weights[1]['b']\n",
    "        \n",
    "        psi = np.dot(s, W0) + b0\n",
    "        x = np.maximum(psi, 0)\n",
    "        dx = (psi > 0).astype(float)\n",
    "\n",
    "        # td_update has the same structure as self.weights, that is an array of dictionaries.\n",
    "        # td_update[0][\"W\"], td_update[0][\"b\"], td_update[1][\"W\"], and td_update[1][\"b\"] have the same shape as \n",
    "        # self.weights[0][\"W\"], self.weights[0][\"b\"], self.weights[1][\"W\"], and self.weights[1][\"b\"] respectively\n",
    "        td_update = [dict() for i in range(len(self.weights))]\n",
    "         \n",
    "        v = delta_mat\n",
    "        td_update[1]['W'] = np.dot(x.T, v) * 1. / s.shape[0]\n",
    "        td_update[1]['b'] = np.sum(v, axis=0, keepdims=True) * 1. / s.shape[0]\n",
    "        \n",
    "        v = np.dot(v, W1.T) * dx\n",
    "        td_update[0]['W'] = np.dot(s.T, v) * 1. / s.shape[0]\n",
    "        td_update[0]['b'] = np.sum(v, axis=0, keepdims=True) * 1. / s.shape[0]\n",
    "                \n",
    "        return td_update\n",
    "    \n",
    "    # Work Required: No. You may wish to read the relevant paper for more information on this weight initialization\n",
    "    # (Exact solutions to the nonlinear dynamics of learning in deep linear neural networks by Saxe, A et al., 2013)\n",
    "    def init_saxe(self, rows, cols):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            rows (int): number of input units for layer.\n",
    "            cols (int): number of output units for layer.\n",
    "        Returns:\n",
    "            NumPy Array consisting of weights for the layer based on the initialization in Saxe et al.\n",
    "        \"\"\"\n",
    "        tensor = self.rand_generator.normal(0, 1, (rows, cols))\n",
    "        if rows < cols:\n",
    "            tensor = tensor.T\n",
    "        tensor, r = np.linalg.qr(tensor)\n",
    "        d = np.diag(r, 0)\n",
    "        ph = np.sign(d)\n",
    "        tensor *= ph\n",
    "\n",
    "        if rows < cols:\n",
    "            tensor = tensor.T\n",
    "        return tensor\n",
    "    \n",
    "    # Work Required: No.\n",
    "    def get_weights(self):\n",
    "        \"\"\"\n",
    "        Returns: \n",
    "            A copy of the current weights of this network.\n",
    "        \"\"\"\n",
    "        return deepcopy(self.weights)\n",
    "    \n",
    "    # Work Required: No.\n",
    "    def set_weights(self, weights):\n",
    "        \"\"\"\n",
    "        Args: \n",
    "            weights (list of dictionaries): Consists of weights that this network will set as its own weights.\n",
    "        \"\"\"\n",
    "        self.weights = deepcopy(weights)"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "61b88349f16b3681a89d0a0b5f7becd6",
     "grade": false,
     "grade_id": "cell6",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the cell below to test your implementation of the `__init__()` function for ActionValueNetwork:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "layer_sizes: (5, 20, 3)\nPassed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for ActionValueNetwork __init__() ## \n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "network_config = {\n",
    "    \"state_dim\": 5,\n",
    "    \"num_hidden_units\": 20,\n",
    "    \"num_actions\": 3\n",
    "}\n",
    "\n",
    "test_network = ActionValueNetwork(network_config)\n",
    "print(\"layer_sizes:\", test_network.layer_sizes)\n",
    "assert(np.allclose(test_network.layer_sizes, np.array([5, 20, 3])))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "d01bed3174f24e2faacde8268f2dd3af",
     "grade": false,
     "grade_id": "cell8",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "**Expected output:**\n",
    "\n",
    "    layer_sizes: [ 5 20  3]"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "5b1d3111fcd72e10d6fae596c1982b11",
     "grade": false,
     "grade_id": "cell9",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 2: Adam Optimizer\n",
    "\n",
    "In this assignment, you will use the Adam algorithm for updating the weights of your action-value network. As you may remember from Course 3 Assignment 2, the Adam algorithm is a more advanced variant of stochastic gradient descent (SGD). The Adam algorithm improves the SGD update with two concepts: adaptive vector stepsizes and momentum. It keeps running estimates of the mean and second moment of the updates, denoted by $\\mathbf{m}$ and $\\mathbf{v}$ respectively:\n",
    "$$\\mathbf{m_t} = \\beta_m \\mathbf{m_{t-1}} + (1 - \\beta_m)g_t \\\\\n",
    "\\mathbf{v_t} = \\beta_v \\mathbf{v_{t-1}} + (1 - \\beta_v)g^2_t\n",
    "$$\n",
    "\n",
    "Here, $\\beta_m$ and $\\beta_v$ are fixed parameters controlling the linear combinations above and $g_t$ is the update at time $t$ (generally the gradients, but here the TD error times the gradients).\n",
    "\n",
    "Given that $\\mathbf{m}$ and $\\mathbf{v}$ are initialized to zero, they are biased toward zero. To get unbiased estimates of the mean and second moment, Adam defines $\\mathbf{\\hat{m}}$ and $\\mathbf{\\hat{v}}$ as:\n",
    "$$ \\mathbf{\\hat{m}_t} = \\frac{\\mathbf{m_t}}{1 - \\beta_m^t} \\\\\n",
    "\\mathbf{\\hat{v}_t} = \\frac{\\mathbf{v_t}}{1 - \\beta_v^t}\n",
    "$$\n",
    "\n",
    "The weights are then updated as follows:\n",
    "$$ \\mathbf{w_t} = \\mathbf{w_{t-1}} + \\frac{\\alpha}{\\sqrt{\\mathbf{\\hat{v}_t}}+\\epsilon} \\mathbf{\\hat{m}_t}\n",
    "$$\n",
    "\n",
    "Here, $\\alpha$ is the step size parameter and $\\epsilon$ is another small parameter to keep the denominator from being zero.\n",
    "\n",
    "In the cell below, you will implement the `__init__()` and `update_weights()` methods for the Adam algorithm. In `__init__()`, you will initialize `self.m` and `self.v`. In `update_weights()`, you will compute new weights given the input weights and an update $g$ (here `td_errors_times_gradients`) according to the equations above."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Work Required: Yes. Fill in code in __init__ and update_weights (~9-11 Lines).\n",
    "class Adam():\n",
    "    # Work Required: Yes. Fill in the initialization for self.m and self.v (~4 Lines).\n",
    "    def __init__(self, layer_sizes, \n",
    "                 optimizer_info):\n",
    "        self.layer_sizes = layer_sizes\n",
    "\n",
    "        # Specify Adam algorithm's hyper parameters\n",
    "        self.step_size = optimizer_info.get(\"step_size\")\n",
    "        self.beta_m = optimizer_info.get(\"beta_m\")\n",
    "        self.beta_v = optimizer_info.get(\"beta_v\")\n",
    "        self.epsilon = optimizer_info.get(\"epsilon\")\n",
    "        \n",
    "        # Initialize Adam algorithm's m and v\n",
    "        self.m = [dict() for i in range(1, len(self.layer_sizes))]\n",
    "        self.v = [dict() for i in range(1, len(self.layer_sizes))]\n",
    "        \n",
    "        for i in range(0, len(self.layer_sizes) - 1):\n",
    "            ### START CODE HERE (~4 Lines)\n",
    "            # Hint: The initialization for m and v should look very much like the initializations of the weights\n",
    "            # except for the fact that initialization here is to zeroes (see description above.)\n",
    "            self.m[i][\"W\"] = np.zeros((self.layer_sizes[i], self.layer_sizes[i+1]))\n",
    "            self.m[i][\"b\"] = np.zeros((1, self.layer_sizes[i+1]))\n",
    "            self.v[i][\"W\"] = np.zeros((self.layer_sizes[i], self.layer_sizes[i+1]))\n",
    "            self.v[i][\"b\"] = np.zeros((1, self.layer_sizes[i+1]))\n",
    "            ### END CODE HERE\n",
    "            \n",
    "        # Notice that to calculate m_hat and v_hat, we use powers of beta_m and beta_v to \n",
    "        # the time step t. We can calculate these powers using an incremental product. At initialization then, \n",
    "        # beta_m_product and beta_v_product should be ...? (Note that timesteps start at 1 and if we were to \n",
    "        # start from 0, the denominator would be 0.)\n",
    "        self.beta_m_product = self.beta_m\n",
    "        self.beta_v_product = self.beta_v\n",
    "    \n",
    "    # Work Required: Yes. Fill in the weight updates (~5-7 lines).\n",
    "    def update_weights(self, weights, td_errors_times_gradients):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            weights (Array of dictionaries): The weights of the neural network.\n",
    "            td_errors_times_gradients (Array of dictionaries): The gradient of the \n",
    "            action-values with respect to the network's weights times the TD-error\n",
    "        Returns:\n",
    "            The updated weights (Array of dictionaries).\n",
    "        \"\"\"\n",
    "        for i in range(len(weights)):\n",
    "            for param in weights[i].keys():\n",
    "                ### START CODE HERE (~5-7 Lines)\n",
    "                # Hint: Follow the equations above. First, you should update m and v and then compute \n",
    "                # m_hat and v_hat. Finally, compute how much the weights should be incremented by.\n",
    "                # self.m[i][param] = None\n",
    "                # self.v[i][param] = None\n",
    "                # m_hat = None\n",
    "                # v_hat = None\n",
    "                self.m[i][param] = self.beta_m*self.m[i][param] + (1 - self.beta_m)*td_errors_times_gradients[i][param]\n",
    "                self.v[i][param] = self.beta_v*self.v[i][param] + (1 - self.beta_v)*np.power(td_errors_times_gradients[i][param], 2)\n",
    "                m_hat = self.m[i][param] / (1 - self.beta_m_product)\n",
    "                v_hat = self.v[i][param] / (1 - self.beta_v_product)\n",
    "                weight_update = self.step_size*(m_hat / (np.sqrt(v_hat) + self.epsilon))\n",
    "                ### END CODE HERE\n",
    "                \n",
    "                weights[i][param] = weights[i][param] + weight_update\n",
    "        # Notice that to calculate m_hat and v_hat, we use powers of beta_m and beta_v to \n",
    "        ### update self.beta_m_product and self.beta_v_product\n",
    "        self.beta_m_product *= self.beta_m\n",
    "        self.beta_v_product *= self.beta_v\n",
    "        \n",
    "        return weights"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "3e71e0d2337c3a571caddde38eed8b8e",
     "grade": false,
     "grade_id": "cell11",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `__init__()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "m[0][\"W\"] shape: (5, 2)\nm[0][\"b\"] shape: (1, 2)\nm[1][\"W\"] shape: (2, 3)\nm[1][\"b\"] shape: (1, 3) \n\nv[0][\"W\"] shape: (5, 2)\nv[0][\"b\"] shape: (1, 2)\nv[1][\"W\"] shape: (2, 3)\nv[1][\"b\"] shape: (1, 3) \n\nPassed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for Adam __init__() ##\n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "network_config = {\"state_dim\": 5,\n",
    "                  \"num_hidden_units\": 2,\n",
    "                  \"num_actions\": 3\n",
    "                 }\n",
    "\n",
    "optimizer_info = {\"step_size\": 0.1,\n",
    "                  \"beta_m\": 0.99,\n",
    "                  \"beta_v\": 0.999,\n",
    "                  \"epsilon\": 0.0001\n",
    "                 }\n",
    "\n",
    "network = ActionValueNetwork(network_config)\n",
    "test_adam = Adam(network.layer_sizes, optimizer_info)\n",
    "\n",
    "print(\"m[0][\\\"W\\\"] shape: {}\".format(test_adam.m[0][\"W\"].shape))\n",
    "print(\"m[0][\\\"b\\\"] shape: {}\".format(test_adam.m[0][\"b\"].shape))\n",
    "print(\"m[1][\\\"W\\\"] shape: {}\".format(test_adam.m[1][\"W\"].shape))\n",
    "print(\"m[1][\\\"b\\\"] shape: {}\".format(test_adam.m[1][\"b\"].shape), \"\\n\")\n",
    "\n",
    "assert(np.allclose(test_adam.m[0][\"W\"].shape, np.array([5, 2])))\n",
    "assert(np.allclose(test_adam.m[0][\"b\"].shape, np.array([1, 2])))\n",
    "assert(np.allclose(test_adam.m[1][\"W\"].shape, np.array([2, 3])))\n",
    "assert(np.allclose(test_adam.m[1][\"b\"].shape, np.array([1, 3])))\n",
    "\n",
    "print(\"v[0][\\\"W\\\"] shape: {}\".format(test_adam.v[0][\"W\"].shape))\n",
    "print(\"v[0][\\\"b\\\"] shape: {}\".format(test_adam.v[0][\"b\"].shape))\n",
    "print(\"v[1][\\\"W\\\"] shape: {}\".format(test_adam.v[1][\"W\"].shape))\n",
    "print(\"v[1][\\\"b\\\"] shape: {}\".format(test_adam.v[1][\"b\"].shape), \"\\n\")\n",
    "\n",
    "assert(np.allclose(test_adam.v[0][\"W\"].shape, np.array([5, 2])))\n",
    "assert(np.allclose(test_adam.v[0][\"b\"].shape, np.array([1, 2])))\n",
    "assert(np.allclose(test_adam.v[1][\"W\"].shape, np.array([2, 3])))\n",
    "assert(np.allclose(test_adam.v[1][\"b\"].shape, np.array([1, 3])))\n",
    "\n",
    "assert(np.all(test_adam.m[0][\"W\"]==0))\n",
    "assert(np.all(test_adam.m[0][\"b\"]==0))\n",
    "assert(np.all(test_adam.m[1][\"W\"]==0))\n",
    "assert(np.all(test_adam.m[1][\"b\"]==0))\n",
    "\n",
    "assert(np.all(test_adam.v[0][\"W\"]==0))\n",
    "assert(np.all(test_adam.v[0][\"b\"]==0))\n",
    "assert(np.all(test_adam.v[1][\"W\"]==0))\n",
    "assert(np.all(test_adam.v[1][\"b\"]==0))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "85dd186066f2dd85643cdc1fa6969a9d",
     "grade": false,
     "grade_id": "cell13",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "**Expected output:**\n",
    "\n",
    "    m[0][\"W\"] shape: (5, 2)\n",
    "    m[0][\"b\"] shape: (1, 2)\n",
    "    m[1][\"W\"] shape: (2, 3)\n",
    "    m[1][\"b\"] shape: (1, 3) \n",
    "\n",
    "    v[0][\"W\"] shape: (5, 2)\n",
    "    v[0][\"b\"] shape: (1, 2)\n",
    "    v[1][\"W\"] shape: (2, 3)\n",
    "    v[1][\"b\"] shape: (1, 3) "
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "959fb7067515ac663cc0b3a6e35ed588",
     "grade": false,
     "grade_id": "cell14",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `update_weights()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "updated_weights[0][\"W\"]\n [[-1.03112528  2.08618453]\n [-0.15531623  0.02412129]\n [-0.76656476 -0.65405898]\n [-0.92569612 -0.24916335]\n [-0.92180119  0.72137957]] \n\nupdated_weights[0][\"b\"]\n [[-0.44392532 -0.69588495]] \n\nupdated_weights[1][\"W\"]\n [[ 0.13962892  0.48820826  0.41311548]\n [ 0.3958054  -0.20738072 -0.47172585]] \n\nupdated_weights[1][\"b\"]\n [[-0.48917533 -0.61934122 -1.48771198]] \n\nPassed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for Adam update_weights() ##\n",
    "\n",
    "network_config = {\"state_dim\": 5,\n",
    "                  \"num_hidden_units\": 2,\n",
    "                  \"num_actions\": 3\n",
    "                 }\n",
    "\n",
    "optimizer_info = {\"step_size\": 0.1,\n",
    "                  \"beta_m\": 0.99,\n",
    "                  \"beta_v\": 0.999,\n",
    "                  \"epsilon\": 0.0001\n",
    "                 }\n",
    "\n",
    "network = ActionValueNetwork(network_config)\n",
    "test_adam = Adam(network.layer_sizes, optimizer_info)\n",
    "\n",
    "rand_generator = np.random.RandomState(0)\n",
    "\n",
    "# Initialize m and v\n",
    "test_adam.m[0][\"W\"] = rand_generator.normal(0, 1, (5, 2))\n",
    "test_adam.m[0][\"b\"] = rand_generator.normal(0, 1, (1, 2))\n",
    "test_adam.m[1][\"W\"] = rand_generator.normal(0, 1, (2, 3))\n",
    "test_adam.m[1][\"b\"] = rand_generator.normal(0, 1, (1, 3))\n",
    "\n",
    "test_adam.v[0][\"W\"] = np.abs(rand_generator.normal(0, 1, (5, 2)))\n",
    "test_adam.v[0][\"b\"] = np.abs(rand_generator.normal(0, 1, (1, 2)))\n",
    "test_adam.v[1][\"W\"] = np.abs(rand_generator.normal(0, 1, (2, 3)))\n",
    "test_adam.v[1][\"b\"] = np.abs(rand_generator.normal(0, 1, (1, 3)))\n",
    "\n",
    "# Specify weights\n",
    "weights = [dict() for i in range(1, len(test_adam.layer_sizes))]\n",
    "weights[0][\"W\"] = rand_generator.normal(0, 1, (5, 2))\n",
    "weights[0][\"b\"] = rand_generator.normal(0, 1, (1, 2))\n",
    "weights[1][\"W\"] = rand_generator.normal(0, 1, (2, 3))\n",
    "weights[1][\"b\"] = rand_generator.normal(0, 1, (1, 3))\n",
    "\n",
    "# Specify g\n",
    "g = [dict() for i in range(1, len(test_adam.layer_sizes))]\n",
    "g[0][\"W\"] = rand_generator.normal(0, 1, (5, 2))\n",
    "g[0][\"b\"] = rand_generator.normal(0, 1, (1, 2))\n",
    "g[1][\"W\"] = rand_generator.normal(0, 1, (2, 3))\n",
    "g[1][\"b\"] = rand_generator.normal(0, 1, (1, 3))\n",
    "\n",
    "# Update weights\n",
    "updated_weights = test_adam.update_weights(weights, g)\n",
    "\n",
    "# updated weights asserts\n",
    "updated_weights_answer = np.load(\"asserts/update_weights.npz\")\n",
    "\n",
    "print(\"updated_weights[0][\\\"W\\\"]\\n\", updated_weights[0][\"W\"], \"\\n\")\n",
    "print(\"updated_weights[0][\\\"b\\\"]\\n\", updated_weights[0][\"b\"], \"\\n\")\n",
    "print(\"updated_weights[1][\\\"W\\\"]\\n\", updated_weights[1][\"W\"], \"\\n\")\n",
    "print(\"updated_weights[1][\\\"b\\\"]\\n\", updated_weights[1][\"b\"], \"\\n\")\n",
    "\n",
    "assert(np.allclose(updated_weights[0][\"W\"], updated_weights_answer[\"W0\"]))\n",
    "assert(np.allclose(updated_weights[0][\"b\"], updated_weights_answer[\"b0\"]))\n",
    "assert(np.allclose(updated_weights[1][\"W\"], updated_weights_answer[\"W1\"]))\n",
    "assert(np.allclose(updated_weights[1][\"b\"], updated_weights_answer[\"b1\"]))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "cd4a53197e8ff77bd6851160ba4967e8",
     "grade": false,
     "grade_id": "cell16",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "**Expected output:**\n",
    "\n",
    "    updated_weights[0][\"W\"]\n",
    "     [[-1.03112528  2.08618453]\n",
    "     [-0.15531623  0.02412129]\n",
    "     [-0.76656476 -0.65405898]\n",
    "     [-0.92569612 -0.24916335]\n",
    "     [-0.92180119  0.72137957]] \n",
    "\n",
    "    updated_weights[0][\"b\"]\n",
    "     [[-0.44392532 -0.69588495]] \n",
    "\n",
    "    updated_weights[1][\"W\"]\n",
    "     [[ 0.13962892  0.48820826  0.41311548]\n",
    "     [ 0.3958054  -0.20738072 -0.47172585]] \n",
    "\n",
    "    updated_weights[1][\"b\"]\n",
    "     [[-0.48917533 -0.61934122 -1.48771198]] "
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "1f0f3e5d39449a6b242db88e6b03e066",
     "grade": false,
     "grade_id": "cell17",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 3: Experience Replay Buffers\n",
    "\n",
    "In Course 3, you implemented agents that update value functions once for each sample. We can use a more efficient approach for updating value functions. You have seen an example of an efficient approach in Course 2 when implementing Dyna. The idea behind Dyna is to learn a model using sampled experience, obtain simulated experience from the model, and improve the value function using the simulated experience.\n",
    "\n",
    "Experience replay is a simple method that can get some of the advantages of Dyna by saving a buffer of experience and using the data stored in the buffer as a model. This view of prior data as a model works because the data represents actual transitions from the underlying MDP. Furthermore, as a side note, this kind of model that is not learned and simply a collection of experience can be called non-parametric as it can be ever-growing as opposed to a parametric model where the transitions are learned to be represented with a fixed set of parameters or weights.\n",
    "\n",
    "We have provided the implementation of the experience replay buffer in the cell below. ReplayBuffer includes two main functions: `append()` and `sample()`. `append()` adds an experience transition to the buffer as an array that includes the state, action, reward, terminal flag (indicating termination of the episode), and next_state. `sample()` gets a batch of experiences from the buffer with size `minibatch_size`.\n",
    "\n",
    "You will use the `append()` and `sample()` functions when implementing the agent."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Do not modify this cell! \n",
    "\n",
    "# Work Required: No. However, do go through the code to ensure your understanding is correct.\n",
    "\n",
    "class ReplayBuffer:\n",
    "    def __init__(self, size, minibatch_size, seed):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            size (integer): The size of the replay buffer.              \n",
    "            minibatch_size (integer): The sample size.\n",
    "            seed (integer): The seed for the random number generator. \n",
    "        \"\"\"\n",
    "        self.buffer = []\n",
    "        self.minibatch_size = minibatch_size\n",
    "        self.rand_generator = np.random.RandomState(seed)\n",
    "        self.max_size = size\n",
    "\n",
    "    def append(self, state, action, reward, terminal, next_state):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            state (Numpy array): The state.              \n",
    "            action (integer): The action.\n",
    "            reward (float): The reward.\n",
    "            terminal (integer): 1 if the next state is a terminal state and 0 otherwise.\n",
    "            next_state (Numpy array): The next state.           \n",
    "        \"\"\"\n",
    "        if len(self.buffer) == self.max_size:\n",
    "            del self.buffer[0]\n",
    "        self.buffer.append([state, action, reward, terminal, next_state])\n",
    "\n",
    "    def sample(self):\n",
    "        \"\"\"\n",
    "        Returns:\n",
    "            A list of transition tuples including state, action, reward, terinal, and next_state\n",
    "        \"\"\"\n",
    "        idxs = self.rand_generator.choice(np.arange(len(self.buffer)), size=self.minibatch_size)\n",
    "        return [self.buffer[idx] for idx in idxs]\n",
    "\n",
    "    def size(self):\n",
    "        return len(self.buffer)"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "3065016279c96b7d6929dc042c2aa5d1",
     "grade": false,
     "grade_id": "cell19",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 4: Softmax Policy\n",
    "\n",
    "In this assignment, you will use a softmax policy. One advantage of a softmax policy is that it explores according to the action-values, meaning that an action with a moderate value has a higher chance of getting selected compared to an action with a lower value. Contrast this with an $\\epsilon$-greedy policy which does not consider the individual action values when choosing an exploratory action in a state and instead chooses randomly when doing so.\n",
    "\n",
    "The probability of selecting each action according to the softmax policy is shown below:\n",
    "$$Pr{(A_t=a | S_t=s)} \\hspace{0.1cm} \\dot{=} \\hspace{0.1cm} \\frac{e^{Q(s, a)/\\tau}}{\\sum_{b \\in A}e^{Q(s, b)/\\tau}}$$\n",
    "where $\\tau$ is the temperature parameter which controls how much the agent focuses on the highest valued actions. The smaller the temperature, the more the agent selects the greedy action. Conversely, when the temperature is high, the agent selects among actions more uniformly random.\n",
    "\n",
    "Given that a softmax policy exponentiates action values, if those values are large, exponentiating them could get very large. To implement the softmax policy in a numerically stable way, we often subtract the maximum action-value from the action-values. If we do so, the probability of selecting each action looks as follows:\n",
    "\n",
    "$$Pr{(A_t=a | S_t=s)} \\hspace{0.1cm} \\dot{=} \\hspace{0.1cm} \\frac{e^{Q(s, a)/\\tau - max_{c}Q(s, c)/\\tau}}{\\sum_{b \\in A}e^{Q(s, b)/\\tau - max_{c}Q(s, c)/\\tau}}$$\n",
    "\n",
    "In the cell below, you will implement the `softmax()` function. In order to do so, you could break the above computation into smaller steps:\n",
    "- compute the preference, $H(a)$, for taking each action by dividing the action-values by the temperature parameter $\\tau$,\n",
    "- subtract the maximum preference across the actions from the preferences to avoid overflow, and,\n",
    "- compute the probability of taking each action."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "def softmax(action_values, tau=1.0):\n",
    "    \"\"\"\n",
    "    Args:\n",
    "        action_values (Numpy array): A 2D array of shape (batch_size, num_actions). \n",
    "                       The action-values computed by an action-value network.              \n",
    "        tau (float): The temperature parameter scalar.\n",
    "    Returns:\n",
    "        A 2D array of shape (batch_size, num_actions). Where each column is a probability distribution over\n",
    "        the actions representing the policy.\n",
    "    \"\"\"\n",
    "    ### START CODE HERE (~2 Lines)\n",
    "    # Compute the preferences by dividing the action-values by the temperature parameter tau\n",
    "    preferences = action_values / tau\n",
    "    # Compute the maximum preference across the actions\n",
    "    max_preference = np.max(preferences, 1)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    \n",
    "    # Reshape max_preference array which has shape [Batch,] to [Batch, 1]. This allows NumPy broadcasting \n",
    "    # when subtracting the maximum preference from the preference of each action.\n",
    "    reshaped_max_preference = max_preference.reshape((-1, 1))\n",
    "    \n",
    "    ### START CODE HERE (~2 Lines)\n",
    "    # Compute the numerator, i.e., the exponential of the preference - the max preference.\n",
    "    exp_preferences = np.exp(preferences - reshaped_max_preference)\n",
    "    # Compute the denominator, i.e., the sum over the numerator along the actions axis.\n",
    "    sum_of_exp_preferences = np.sum(exp_preferences, 1)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    \n",
    "    # Reshape sum_of_exp_preferences array which has shape [Batch,] to [Batch, 1] to  allow for NumPy broadcasting \n",
    "    # when dividing the numerator by the denominator.\n",
    "    reshaped_sum_of_exp_preferences = sum_of_exp_preferences.reshape((-1, 1))\n",
    "    \n",
    "    ### START CODE HERE (~1 Lines)\n",
    "    # Compute the action probabilities according to the equation in the previous cell.\n",
    "    action_probs = exp_preferences / reshaped_sum_of_exp_preferences\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    \n",
    "    # squeeze() removes any singleton dimensions. It is used here because this function is used in the \n",
    "    # agent policy when selecting an action (for which the batch dimension is 1.) As np.random.choice is used in \n",
    "    # the agent policy and it expects 1D arrays, we need to remove this singleton batch dimension.\n",
    "    action_probs = action_probs.squeeze()\n",
    "    return action_probs"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "78e00022f5e617732b6527f3523724e0",
     "grade": false,
     "grade_id": "cell21",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the cell below to test your implementation of the `softmax()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "action_probs [[0.25849645 0.01689625 0.05374514 0.67086216]\n [0.84699852 0.00286345 0.13520063 0.01493741]]\nPassed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for softmax() ##\n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "rand_generator = np.random.RandomState(0)\n",
    "action_values = rand_generator.normal(0, 1, (2, 4))\n",
    "tau = 0.5\n",
    "\n",
    "action_probs = softmax(action_values, tau)\n",
    "print(\"action_probs\", action_probs)\n",
    "\n",
    "assert(np.allclose(action_probs, np.array([\n",
    "    [0.25849645, 0.01689625, 0.05374514, 0.67086216],\n",
    "    [0.84699852, 0.00286345, 0.13520063, 0.01493741]\n",
    "])))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "a551ed093fa021ec1e222fb06539a63f",
     "grade": false,
     "grade_id": "cell23",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "**Expected output:**\n",
    "\n",
    "    action_probs [[0.25849645 0.01689625 0.05374514 0.67086216]\n",
    "     [0.84699852 0.00286345 0.13520063 0.01493741]]"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "e66eb8fa7dda54905f0b5a45aea2fc88",
     "grade": false,
     "grade_id": "cell24",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 5: Putting the pieces together\n",
    "\n",
    "In this section, you will combine components from the previous sections to write up an RL-Glue Agent. The main component that you will implement is the action-value network updates with experience sampled from the experience replay buffer.\n",
    "\n",
    "At time $t$, we have an action-value function represented as a neural network, say $Q_t$. We want to update our action-value function and get a new one we can use at the next timestep. We will get this $Q_{t+1}$ using multiple replay steps that each result in an intermediate action-value function $Q_{t+1}^{i}$ where $i$ indexes which replay step we are at.\n",
    "\n",
    "In each replay step, we sample a batch of experiences from the replay buffer and compute a minibatch Expected-SARSA update. Across these N replay steps, we will use the current \"un-updated\" action-value network at time $t$, $Q_t$, for computing the action-values of the next-states. This contrasts using the most recent action-values from the last replay step $Q_{t+1}^{i}$. We make this choice to have targets that are stable across replay steps. Here is the pseudocode for performing the updates:\n",
    "\n",
    "$$\n",
    "\\begin{align}\n",
    "& Q_t \\leftarrow \\text{action-value network at timestep t (current action-value network)}\\\\\n",
    "& \\text{Initialize } Q_{t+1}^1 \\leftarrow Q_t\\\\\n",
    "& \\text{For } i \\text{ in } [1, ..., N] \\text{ (i.e. N} \\text{  replay steps)}:\\\\\n",
    "& \\hspace{1cm} s, a, r, t, s'\n",
    "\\leftarrow \\text{Sample batch of experiences from experience replay buffer} \\\\\n",
    "& \\hspace{1cm} \\text{Do Expected Sarsa update with } Q_t: Q_{t+1}^{i+1}(s, a) \\leftarrow Q_{t+1}^{i}(s, a) + \\alpha \\cdot \\left[r + \\gamma \\left(\\sum_{b} \\pi(b | s') Q_t(s', b)\\right) - Q_{t+1}^{i}(s, a)\\right]\\\\\n",
    "& \\hspace{1.5cm} \\text{ making sure to add the } \\gamma \\left(\\sum_{b} \\pi(b | s') Q_t(s', b)\\right) \\text{ for non-terminal transitions only.} \\\\\n",
    "& \\text{After N replay steps, we set } Q_{t+1}^{N} \\text{ as } Q_{t+1} \\text{ and have a new } Q_{t+1} \\text{for time step } t + 1 \\text{ that we will fix in the next set of updates. }\n",
    "\\end{align}\n",
    "$$\n",
    "\n",
    "As you can see in the pseudocode, after sampling a batch of experiences, we do many computations. The basic idea however is that we are looking to compute a form of a TD error. In order to so, we can take the following steps:\n",
    "- compute the action-values for the next states using the action-value network $Q_{t}$,\n",
    "- compute the policy $\\pi(b | s')$ induced by the action-values $Q_{t}$ (using the softmax function you implemented before),\n",
    "- compute the Expected sarsa targets $r + \\gamma \\left(\\sum_{b} \\pi(b | s') Q_t(s', b)\\right)$,\n",
    "- compute the action-values for the current states using the latest $Q_{t + 1}$, and,\n",
    "- compute the TD-errors with the Expected Sarsa targets.\n",
    " \n",
    "For the third step above, you can start by computing $\\pi(b | s') Q_t(s', b)$ followed by summation to get $\\hat{v}_\\pi(s') = \\left(\\sum_{b} \\pi(b | s') Q_t(s', b)\\right)$. $\\hat{v}_\\pi(s')$ is an estimate of the value of the next state. Note for terminal next states, $\\hat{v}_\\pi(s') = 0$. Finally, we add the rewards to the discount times $\\hat{v}_\\pi(s')$.\n",
    "\n",
    "You will implement these steps in the `get_td_error()` function below which given a batch of experiences (including states, next_states, actions, rewards, terminals), fixed action-value network (current_q), and action-value network (network), computes the TD error in the form of a 1D array of size batch_size."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Work Required: Yes. Fill in code in get_td_error (~9 Lines).\n",
    "def get_td_error(states, next_states, actions, rewards, discount, terminals, network, current_q, tau):\n",
    "    \"\"\"\n",
    "    Args:\n",
    "        states (Numpy array): The batch of states with the shape (batch_size, state_dim).\n",
    "        next_states (Numpy array): The batch of next states with the shape (batch_size, state_dim).\n",
    "        actions (Numpy array): The batch of actions with the shape (batch_size,).\n",
    "        rewards (Numpy array): The batch of rewards with the shape (batch_size,).\n",
    "        discount (float): The discount factor.\n",
    "        terminals (Numpy array): The batch of terminals with the shape (batch_size,).\n",
    "        network (ActionValueNetwork): The latest state of the network that is getting replay updates.\n",
    "        current_q (ActionValueNetwork): The fixed network used for computing the targets, \n",
    "                                        and particularly, the action-values at the next-states.\n",
    "    Returns:\n",
    "        The TD errors (Numpy array) for actions taken, of shape (batch_size,)\n",
    "    \"\"\"\n",
    "    \n",
    "    # Note: Here network is the latest state of the network that is getting replay updates. In other words, \n",
    "    # the network represents Q_{t+1}^{i} whereas current_q represents Q_t, the fixed network used for computing the \n",
    "    # targets, and particularly, the action-values at the next-states.\n",
    "    \n",
    "    # Compute action values at next states using current_q network\n",
    "    # Note that q_next_mat is a 2D array of shape (batch_size, num_actions)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    q_next_mat = current_q.get_action_values(next_states)  # (batch_size, num_actions)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Compute policy at next state by passing the action-values in q_next_mat to softmax()\n",
    "    # Note that probs_mat is a 2D array of shape (batch_size, num_actions)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    probs_mat = softmax(q_next_mat, tau)  # (batch_size, num_actions)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Compute the estimate of the next state value, v_next_vec.\n",
    "    # Hint: sum the action-values for the next_states weighted by the policy, probs_mat. Then, multiply by\n",
    "    # (1 - terminals) to make sure v_next_vec is zero for terminal next states.\n",
    "    # Note that v_next_vec is a 1D array of shape (batch_size,)\n",
    "    \n",
    "    ### START CODE HERE (~3 Lines)\n",
    "    v_next_vec = np.sum(np.multiply(q_next_mat, probs_mat), 1)*(1 - terminals)  # (1, batch_size)\n",
    "    v_next_vec = v_next_vec.reshape(-1, )  # (batch_size,)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Compute Expected Sarsa target\n",
    "    # Note that target_vec is a 1D array of shape (batch_size,)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    target_vec = rewards + discount*v_next_vec  # (batch_size,)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Compute action values at the current states for all actions using network\n",
    "    # Note that q_mat is a 2D array of shape (batch_size, num_actions)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    q_mat = network.get_action_values(states)  # (batch_size, num_actions)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Batch Indices is an array from 0 to the batch size - 1. \n",
    "    batch_indices = np.arange(q_mat.shape[0])\n",
    "\n",
    "    # Compute q_vec by selecting q(s, a) from q_mat for taken actions\n",
    "    # Use batch_indices as the index for the first dimension of q_mat\n",
    "    # Note that q_vec is a 1D array of shape (batch_size)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    q_vec = q_mat[batch_indices, actions]\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Compute TD errors for actions taken\n",
    "    # Note that delta_vec is a 1D array of shape (batch_size)\n",
    "    \n",
    "    ### START CODE HERE (~1 Line)\n",
    "    delta_vec = target_vec - q_vec\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    return delta_vec"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "7efa3ae7896324abe10eff83fa20c3a1",
     "grade": false,
     "grade_id": "cell26",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `get_td_error()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for get_td_error() ##\n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "data = np.load(\"asserts/get_td_error_1.npz\", allow_pickle=True)\n",
    "\n",
    "states = data[\"states\"]\n",
    "next_states = data[\"next_states\"]\n",
    "actions = data[\"actions\"]\n",
    "rewards = data[\"rewards\"]\n",
    "discount = data[\"discount\"]\n",
    "terminals = data[\"terminals\"]\n",
    "tau = 0.001\n",
    "\n",
    "network_config = {\"state_dim\": 8,\n",
    "                  \"num_hidden_units\": 512,\n",
    "                  \"num_actions\": 4\n",
    "                  }\n",
    "\n",
    "network = ActionValueNetwork(network_config)\n",
    "network.set_weights(data[\"network_weights\"])\n",
    "\n",
    "current_q = ActionValueNetwork(network_config)\n",
    "current_q.set_weights(data[\"current_q_weights\"])\n",
    "\n",
    "delta_vec = get_td_error(states, next_states, actions, rewards, discount, terminals, network, current_q, tau)\n",
    "answer_delta_vec = data[\"delta_vec\"]\n",
    "\n",
    "assert(np.allclose(delta_vec, answer_delta_vec))\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "0ea7393d98a54b3f73abf2177a499e83",
     "grade": false,
     "grade_id": "cell28",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Now that you implemented the `get_td_error()` function, you can use it to implement the `optimize_network()` function. In this function, you will:\n",
    "- get the TD-errors vector from `get_td_error()`,\n",
    "- make the TD-errors into a matrix using zeroes for actions not taken in the transitions,\n",
    "- pass the TD-errors matrix to the `get_TD_update()` function of network to calculate the gradients times TD errors, and,\n",
    "- perform an ADAM optimizer step."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Work Required: Yes. Fill in code in optimize_network (~2 Lines).\n",
    "def optimize_network(experiences, discount, optimizer, network, current_q, tau):\n",
    "    \"\"\"\n",
    "    Args:\n",
    "        experiences (Numpy array): The batch of experiences including the states, actions, \n",
    "                                   rewards, terminals, and next_states.\n",
    "        discount (float): The discount factor.\n",
    "        network (ActionValueNetwork): The latest state of the network that is getting replay updates.\n",
    "        current_q (ActionValueNetwork): The fixed network used for computing the targets, \n",
    "                                        and particularly, the action-values at the next-states.\n",
    "    \"\"\"\n",
    "    \n",
    "    # Get states, action, rewards, terminals, and next_states from experiences\n",
    "    states, actions, rewards, terminals, next_states = map(list, zip(*experiences))\n",
    "    states = np.concatenate(states)\n",
    "    next_states = np.concatenate(next_states)\n",
    "    rewards = np.array(rewards)\n",
    "    terminals = np.array(terminals)\n",
    "    batch_size = states.shape[0]\n",
    "\n",
    "    # Compute TD error using the get_td_error function\n",
    "    # Note that q_vec is a 1D array of shape (batch_size)\n",
    "    delta_vec = get_td_error(states, next_states, actions, rewards, discount, terminals, network, current_q, tau)\n",
    "\n",
    "    # Batch Indices is an array from 0 to the batch_size - 1. \n",
    "    batch_indices = np.arange(batch_size)\n",
    "\n",
    "    # Make a td error matrix of shape (batch_size, num_actions)\n",
    "    # delta_mat has non-zero value only for actions taken\n",
    "    delta_mat = np.zeros((batch_size, network.num_actions))\n",
    "    delta_mat[batch_indices, actions] = delta_vec\n",
    "\n",
    "    # Pass delta_mat to compute the TD errors times the gradients of the network's weights from back-propagation\n",
    "    \n",
    "    ### START CODE HERE\n",
    "    td_update = network.get_TD_update(states, delta_mat)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    # Pass network.get_weights and the td_update to the optimizer to get updated weights\n",
    "    ### START CODE HERE\n",
    "    weights = optimizer.update_weights(network.get_weights(), td_update)\n",
    "    ### END CODE HERE\n",
    "    \n",
    "    network.set_weights(weights)"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "3889eb0e75da92c21d961ec401627989",
     "grade": false,
     "grade_id": "cell30",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `optimize_network()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for optimize_network() ##\n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "input_data = np.load(\"asserts/optimize_network_input_1.npz\", allow_pickle=True)\n",
    "\n",
    "experiences = list(input_data[\"experiences\"])\n",
    "discount = input_data[\"discount\"]\n",
    "tau = 0.001\n",
    "\n",
    "network_config = {\"state_dim\": 8,\n",
    "                  \"num_hidden_units\": 512,\n",
    "                  \"num_actions\": 4\n",
    "                  }\n",
    "\n",
    "network = ActionValueNetwork(network_config)\n",
    "network.set_weights(input_data[\"network_weights\"])\n",
    "\n",
    "current_q = ActionValueNetwork(network_config)\n",
    "current_q.set_weights(input_data[\"current_q_weights\"])\n",
    "\n",
    "optimizer_config = {'step_size': 3e-5, \n",
    "                    'beta_m': 0.9, \n",
    "                    'beta_v': 0.999,\n",
    "                    'epsilon': 1e-8\n",
    "                   }\n",
    "optimizer = Adam(network.layer_sizes, optimizer_config)\n",
    "optimizer.m = input_data[\"optimizer_m\"]\n",
    "optimizer.v = input_data[\"optimizer_v\"]\n",
    "optimizer.beta_m_product = input_data[\"optimizer_beta_m_product\"]\n",
    "optimizer.beta_v_product = input_data[\"optimizer_beta_v_product\"]\n",
    "\n",
    "optimize_network(experiences, discount, optimizer, network, current_q, tau)\n",
    "updated_weights = network.get_weights()\n",
    "\n",
    "output_data = np.load(\"asserts/optimize_network_output_1.npz\", allow_pickle=True)\n",
    "answer_updated_weights = output_data[\"updated_weights\"]\n",
    "\n",
    "assert(np.allclose(updated_weights[0][\"W\"], answer_updated_weights[0][\"W\"]))\n",
    "assert(np.allclose(updated_weights[0][\"b\"], answer_updated_weights[0][\"b\"]))\n",
    "assert(np.allclose(updated_weights[1][\"W\"], answer_updated_weights[1][\"W\"]))\n",
    "assert(np.allclose(updated_weights[1][\"b\"], answer_updated_weights[1][\"b\"]))\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "5b83a775d274b1dc0d9ee3e3945fcef8",
     "grade": false,
     "grade_id": "cell32",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Now that you implemented the `optimize_network()` function, you can implement the agent. In the cell below, you will fill the `agent_step()` and `agent_end()` functions. You should:\n",
    "- select an action (only in `agent_step()`),\n",
    "- add transitions (consisting of the state, action, reward, terminal, and next state) to the replay buffer, and,\n",
    "- update the weights of the neural network by doing multiple replay steps and calling the `optimize_network()` function that you implemented above."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Work Required: Yes. Fill in code in agent_step and agent_end (~7 Lines).\n",
    "class Agent(BaseAgent):\n",
    "    def __init__(self):\n",
    "        self.name = \"expected_sarsa_agent\"\n",
    "        \n",
    "    # Work Required: No.\n",
    "    def agent_init(self, agent_config):\n",
    "        \"\"\"Setup for the agent called when the experiment first starts.\n",
    "\n",
    "        Set parameters needed to setup the agent.\n",
    "\n",
    "        Assume agent_config dict contains:\n",
    "        {\n",
    "            network_config: dictionary,\n",
    "            optimizer_config: dictionary,\n",
    "            replay_buffer_size: integer,\n",
    "            minibatch_sz: integer, \n",
    "            num_replay_updates_per_step: float\n",
    "            discount_factor: float,\n",
    "        }\n",
    "        \"\"\"\n",
    "        self.replay_buffer = ReplayBuffer(agent_config['replay_buffer_size'], \n",
    "                                          agent_config['minibatch_sz'], agent_config.get(\"seed\"))\n",
    "        self.network = ActionValueNetwork(agent_config['network_config'])\n",
    "        self.optimizer = Adam(self.network.layer_sizes, agent_config[\"optimizer_config\"])\n",
    "        self.num_actions = agent_config['network_config']['num_actions']\n",
    "        self.num_replay = agent_config['num_replay_updates_per_step']\n",
    "        self.discount = agent_config['gamma']\n",
    "        self.tau = agent_config['tau']\n",
    "        \n",
    "        self.rand_generator = np.random.RandomState(agent_config.get(\"seed\"))\n",
    "        \n",
    "        self.last_state = None\n",
    "        self.last_action = None\n",
    "        \n",
    "        self.sum_rewards = 0\n",
    "        self.episode_steps = 0\n",
    "\n",
    "    # Work Required: No.\n",
    "    def policy(self, state):\n",
    "        \"\"\"\n",
    "        Args:\n",
    "            state (Numpy array): the state.\n",
    "        Returns:\n",
    "            the action. \n",
    "        \"\"\"\n",
    "        action_values = self.network.get_action_values(state)\n",
    "        probs_batch = softmax(action_values, self.tau)\n",
    "        action = self.rand_generator.choice(self.num_actions, p=probs_batch.squeeze())\n",
    "        return action\n",
    "\n",
    "    # Work Required: No.\n",
    "    def agent_start(self, state):\n",
    "        \"\"\"The first method called when the experiment starts, called after\n",
    "        the environment starts.\n",
    "        Args:\n",
    "            state (Numpy array): the state from the\n",
    "                environment's evn_start function.\n",
    "        Returns:\n",
    "            The first action the agent takes.\n",
    "        \"\"\"\n",
    "        self.sum_rewards = 0\n",
    "        self.episode_steps = 0\n",
    "        self.last_state = np.array([state])\n",
    "        self.last_action = self.policy(self.last_state)\n",
    "        return self.last_action\n",
    "\n",
    "    # Work Required: Yes. Fill in the action selection, replay-buffer update, \n",
    "    # weights update using optimize_network, and updating last_state and last_action (~5 lines).\n",
    "    def agent_step(self, reward, state):\n",
    "        \"\"\"A step taken by the agent.\n",
    "        Args:\n",
    "            reward (float): the reward received for taking the last action taken\n",
    "            state (Numpy array): the state from the\n",
    "                environment's step based, where the agent ended up after the\n",
    "                last step\n",
    "        Returns:\n",
    "            The action the agent is taking.\n",
    "        \"\"\"\n",
    "        \n",
    "        self.sum_rewards += reward\n",
    "        self.episode_steps += 1\n",
    "\n",
    "        # Make state an array of shape (1, state_dim) to add a batch dimension and\n",
    "        # to later match the get_action_values() and get_TD_update() functions\n",
    "        state = np.array([state])\n",
    "\n",
    "        # Select action\n",
    "        ### START CODE HERE (~1 Line)\n",
    "        action = self.policy(state)\n",
    "        ### END CODE HERE\n",
    "        \n",
    "        # Append new experience to replay buffer\n",
    "        # Note: look at the replay_buffer append function for the order of arguments\n",
    "\n",
    "        ### START CODE HERE (~1 Line)\n",
    "        self.replay_buffer.append(self.last_state, self.last_action, reward, False, state)\n",
    "        ### END CODE HERE\n",
    "        \n",
    "        # Perform replay steps:\n",
    "        if self.replay_buffer.size() > self.replay_buffer.minibatch_size:\n",
    "            current_q = deepcopy(self.network)\n",
    "            for _ in range(self.num_replay):\n",
    "                \n",
    "                # Get sample experiences from the replay buffer\n",
    "                experiences = self.replay_buffer.sample()\n",
    "                \n",
    "                # Call optimize_network to update the weights of the network (~1 Line)\n",
    "                ### START CODE HERE\n",
    "                optimize_network(experiences, self.discount, self.optimizer, self.network, current_q, self.tau)\n",
    "                ### END CODE HERE\n",
    "                \n",
    "        # Update the last state and last action.\n",
    "        ### START CODE HERE (~2 Lines)\n",
    "        self.last_state = state\n",
    "        self.last_action = action\n",
    "        ### END CODE HERE\n",
    "        \n",
    "        return action\n",
    "\n",
    "    # Work Required: Yes. Fill in the replay-buffer update and\n",
    "    # update of the weights using optimize_network (~2 lines).\n",
    "    def agent_end(self, reward):\n",
    "        \"\"\"Run when the agent terminates.\n",
    "        Args:\n",
    "            reward (float): the reward the agent received for entering the\n",
    "                terminal state.\n",
    "        \"\"\"\n",
    "        self.sum_rewards += reward\n",
    "        self.episode_steps += 1\n",
    "        \n",
    "        # Set terminal state to an array of zeros\n",
    "        state = np.zeros_like(self.last_state)\n",
    "\n",
    "        # Append new experience to replay buffer\n",
    "        # Note: look at the replay_buffer append function for the order of arguments\n",
    "        \n",
    "        ### START CODE HERE (~1 Line)\n",
    "        self.replay_buffer.append(self.last_state, self.last_action, reward, True, state)\n",
    "        ### END CODE HERE\n",
    "        \n",
    "        # Perform replay steps:\n",
    "        if self.replay_buffer.size() > self.replay_buffer.minibatch_size:\n",
    "            current_q = deepcopy(self.network)\n",
    "            for _ in range(self.num_replay):\n",
    "                \n",
    "                # Get sample experiences from the replay buffer\n",
    "                experiences = self.replay_buffer.sample()\n",
    "                \n",
    "                # Call optimize_network to update the weights of the network\n",
    "                ### START CODE HERE (~1 Line)\n",
    "                optimize_network(experiences, self.discount, self.optimizer, self.network, current_q, self.tau)\n",
    "                ### END CODE HERE\n",
    "                \n",
    "        \n",
    "    def agent_message(self, message):\n",
    "        if message == \"get_sum_reward\":\n",
    "            return self.sum_rewards\n",
    "        else:\n",
    "            raise Exception(\"Unrecognized Message!\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "dab79d391aef354cd1fc454690c534d3",
     "grade": false,
     "grade_id": "cell34",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `agent_step()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for agent_step() ## \n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "agent_info = {\n",
    "             'network_config': {\n",
    "                 'state_dim': 8,\n",
    "                 'num_hidden_units': 256,\n",
    "                 'num_hidden_layers': 1,\n",
    "                 'num_actions': 4\n",
    "             },\n",
    "             'optimizer_config': {\n",
    "                 'step_size': 3e-5, \n",
    "                 'beta_m': 0.9, \n",
    "                 'beta_v': 0.999,\n",
    "                 'epsilon': 1e-8\n",
    "             },\n",
    "             'replay_buffer_size': 32,\n",
    "             'minibatch_sz': 32,\n",
    "             'num_replay_updates_per_step': 4,\n",
    "             'gamma': 0.99,\n",
    "             'tau': 1000.0,\n",
    "             'seed': 0}\n",
    "\n",
    "# Initialize agent\n",
    "agent = Agent()\n",
    "agent.agent_init(agent_info)\n",
    "\n",
    "# load agent network, optimizer, replay_buffer from the agent_input_1.npz file\n",
    "input_data = np.load(\"asserts/agent_input_1.npz\", allow_pickle=True)\n",
    "agent.network.set_weights(input_data[\"network_weights\"])\n",
    "agent.optimizer.m = input_data[\"optimizer_m\"]\n",
    "agent.optimizer.v = input_data[\"optimizer_v\"]\n",
    "agent.optimizer.beta_m_product = input_data[\"optimizer_beta_m_product\"]\n",
    "agent.optimizer.beta_v_product = input_data[\"optimizer_beta_v_product\"]\n",
    "agent.replay_buffer.rand_generator.seed(int(input_data[\"replay_buffer_seed\"]))\n",
    "for experience in input_data[\"replay_buffer\"]:\n",
    "    agent.replay_buffer.buffer.append(experience)\n",
    "\n",
    "# Perform agent_step multiple times\n",
    "last_state_array = input_data[\"last_state_array\"]\n",
    "last_action_array = input_data[\"last_action_array\"]\n",
    "state_array = input_data[\"state_array\"]\n",
    "reward_array = input_data[\"reward_array\"]\n",
    "\n",
    "for i in range(5):\n",
    "    agent.last_state = last_state_array[i]\n",
    "    agent.last_action = last_action_array[i]\n",
    "    state = state_array[i]\n",
    "    reward = reward_array[i]\n",
    "    \n",
    "    agent.agent_step(reward, state)\n",
    "    \n",
    "    # Load expected values for last_state, last_action, weights, and replay_buffer \n",
    "    output_data = np.load(\"asserts/agent_step_output_{}.npz\".format(i), allow_pickle=True)\n",
    "    answer_last_state = output_data[\"last_state\"]\n",
    "    answer_last_action = output_data[\"last_action\"]\n",
    "    answer_updated_weights = output_data[\"updated_weights\"]\n",
    "    answer_replay_buffer = output_data[\"replay_buffer\"]\n",
    "\n",
    "    # Asserts for last_state and last_action\n",
    "    assert(np.allclose(answer_last_state, agent.last_state))\n",
    "    assert(np.allclose(answer_last_action, agent.last_action))\n",
    "\n",
    "    # Asserts for replay_buffer \n",
    "    for i in range(answer_replay_buffer.shape[0]):\n",
    "        for j in range(answer_replay_buffer.shape[1]):\n",
    "            assert(np.allclose(np.asarray(agent.replay_buffer.buffer)[i, j], answer_replay_buffer[i, j]))\n",
    "\n",
    "    # Asserts for network.weights\n",
    "    assert(np.allclose(agent.network.weights[0][\"W\"], answer_updated_weights[0][\"W\"]))\n",
    "    assert(np.allclose(agent.network.weights[0][\"b\"], answer_updated_weights[0][\"b\"]))\n",
    "    assert(np.allclose(agent.network.weights[1][\"W\"], answer_updated_weights[1][\"W\"]))\n",
    "    assert(np.allclose(agent.network.weights[1][\"b\"], answer_updated_weights[1][\"b\"]))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "31468a23f769e117fd042f1be96c3cff",
     "grade": false,
     "grade_id": "cell36",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the following code to test your implementation of the `agent_end()` function:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": "Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\n"
    }
   ],
   "source": [
    "# Do not modify this cell!\n",
    "\n",
    "## Test Code for agent_end() ## \n",
    "\n",
    "# NOTE: The test below is limited in scope. Additional tests are used in the autograder, so it is recommended \n",
    "# to test your implementations more carefully for correctness.\n",
    "\n",
    "agent_info = {\n",
    "             'network_config': {\n",
    "                 'state_dim': 8,\n",
    "                 'num_hidden_units': 256,\n",
    "                 'num_hidden_layers': 1,\n",
    "                 'num_actions': 4\n",
    "             },\n",
    "             'optimizer_config': {\n",
    "                 'step_size': 3e-5, \n",
    "                 'beta_m': 0.9, \n",
    "                 'beta_v': 0.999,\n",
    "                 'epsilon': 1e-8\n",
    "             },\n",
    "             'replay_buffer_size': 32,\n",
    "             'minibatch_sz': 32,\n",
    "             'num_replay_updates_per_step': 4,\n",
    "             'gamma': 0.99,\n",
    "             'tau': 1000,\n",
    "             'seed': 0\n",
    "             }\n",
    "\n",
    "# Initialize agent\n",
    "agent = Agent()\n",
    "agent.agent_init(agent_info)\n",
    "\n",
    "# load agent network, optimizer, replay_buffer from the agent_input_1.npz file\n",
    "input_data = np.load(\"asserts/agent_input_1.npz\", allow_pickle=True)\n",
    "agent.network.set_weights(input_data[\"network_weights\"])\n",
    "agent.optimizer.m = input_data[\"optimizer_m\"]\n",
    "agent.optimizer.v = input_data[\"optimizer_v\"]\n",
    "agent.optimizer.beta_m_product = input_data[\"optimizer_beta_m_product\"]\n",
    "agent.optimizer.beta_v_product = input_data[\"optimizer_beta_v_product\"]\n",
    "agent.replay_buffer.rand_generator.seed(int(input_data[\"replay_buffer_seed\"]))\n",
    "for experience in input_data[\"replay_buffer\"]:\n",
    "    agent.replay_buffer.buffer.append(experience)\n",
    "\n",
    "# Perform agent_step multiple times\n",
    "last_state_array = input_data[\"last_state_array\"]\n",
    "last_action_array = input_data[\"last_action_array\"]\n",
    "state_array = input_data[\"state_array\"]\n",
    "reward_array = input_data[\"reward_array\"]\n",
    "\n",
    "for i in range(5):\n",
    "    agent.last_state = last_state_array[i]\n",
    "    agent.last_action = last_action_array[i]\n",
    "    reward = reward_array[i]\n",
    "    \n",
    "    agent.agent_end(reward)\n",
    "\n",
    "    # Load expected values for last_state, last_action, weights, and replay_buffer \n",
    "    output_data = np.load(\"asserts/agent_end_output_{}.npz\".format(i), allow_pickle=True)\n",
    "    answer_updated_weights = output_data[\"updated_weights\"]\n",
    "    answer_replay_buffer = output_data[\"replay_buffer\"]\n",
    "\n",
    "    # Asserts for replay_buffer \n",
    "    for i in range(answer_replay_buffer.shape[0]):\n",
    "        for j in range(answer_replay_buffer.shape[1]):\n",
    "            assert(np.allclose(np.asarray(agent.replay_buffer.buffer)[i, j], answer_replay_buffer[i, j]))\n",
    "\n",
    "    # Asserts for network.weights\n",
    "    assert(np.allclose(agent.network.weights[0][\"W\"], answer_updated_weights[0][\"W\"]))\n",
    "    assert(np.allclose(agent.network.weights[0][\"b\"], answer_updated_weights[0][\"b\"]))\n",
    "    assert(np.allclose(agent.network.weights[1][\"W\"], answer_updated_weights[1][\"W\"]))\n",
    "    assert(np.allclose(agent.network.weights[1][\"b\"], answer_updated_weights[1][\"b\"]))\n",
    "\n",
    "print(\"Passed the asserts! (Note: These are however limited in scope, additional testing is encouraged.)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "dcbf680f9de1cca0814b50b4a728c47b",
     "grade": false,
     "grade_id": "cell38",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "## Section 6: Run Experiment\n",
    "\n",
    "Now that you implemented the agent, we can use it to run an experiment on the Lunar Lander problem. We will plot the learning curve of the agent to visualize learning progress. To plot the learning curve, we use the sum of rewards in an episode as the performance measure. We have provided for you the experiment/plot code in the cell below which you can go ahead and run. Note that running the cell below has taken approximately 10 minutes in prior testing."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": "100%|██████████| 300/300 [16:39<00:00,  3.33s/it]\n"
    }
   ],
   "source": [
    "def run_experiment(environment, agent, environment_parameters, agent_parameters, experiment_parameters):\n",
    "    \n",
    "    rl_glue = RLGlue(environment, agent)\n",
    "        \n",
    "    # save sum of reward at the end of each episode\n",
    "    agent_sum_reward = np.zeros((experiment_parameters[\"num_runs\"], \n",
    "                                 experiment_parameters[\"num_episodes\"]))\n",
    "\n",
    "    env_info = {}\n",
    "\n",
    "    agent_info = agent_parameters\n",
    "\n",
    "    # one agent setting\n",
    "    for run in range(1, experiment_parameters[\"num_runs\"]+1):\n",
    "        agent_info[\"seed\"] = run\n",
    "        agent_info[\"network_config\"][\"seed\"] = run\n",
    "        env_info[\"seed\"] = run\n",
    "\n",
    "        rl_glue.rl_init(agent_info, env_info)\n",
    "        \n",
    "        for episode in tqdm(range(1, experiment_parameters[\"num_episodes\"]+1)):\n",
    "            # run episode\n",
    "            rl_glue.rl_episode(experiment_parameters[\"timeout\"])\n",
    "            \n",
    "            episode_reward = rl_glue.rl_agent_message(\"get_sum_reward\")\n",
    "            agent_sum_reward[run - 1, episode - 1] = episode_reward\n",
    "    save_name = \"{}\".format(rl_glue.agent.name)\n",
    "    if not os.path.exists('results'):\n",
    "        os.makedirs('results')\n",
    "    np.save(\"results/sum_reward_{}\".format(save_name), agent_sum_reward)\n",
    "    shutil.make_archive('results', 'zip', 'results')\n",
    "\n",
    "# Run Experiment\n",
    "\n",
    "# Experiment parameters\n",
    "experiment_parameters = {\n",
    "    \"num_runs\" : 1,\n",
    "    \"num_episodes\" : 300,\n",
    "    # OpenAI Gym environments allow for a timestep limit timeout, causing episodes to end after \n",
    "    # some number of timesteps. Here we use the default of 1000.\n",
    "    \"timeout\" : 1000\n",
    "}\n",
    "\n",
    "# Environment parameters\n",
    "environment_parameters = {}\n",
    "\n",
    "current_env = LunarLanderEnvironment\n",
    "\n",
    "# Agent parameters\n",
    "agent_parameters = {\n",
    "    'network_config': {\n",
    "        'state_dim': 8,\n",
    "        'num_hidden_units': 256,\n",
    "        'num_actions': 4\n",
    "    },\n",
    "    'optimizer_config': {\n",
    "        'step_size': 1e-3,\n",
    "        'beta_m': 0.9, \n",
    "        'beta_v': 0.999,\n",
    "        'epsilon': 1e-8\n",
    "    },\n",
    "    'replay_buffer_size': 50000,\n",
    "    'minibatch_sz': 8,\n",
    "    'num_replay_updates_per_step': 4,\n",
    "    'gamma': 0.99,\n",
    "    'tau': 0.001\n",
    "}\n",
    "current_agent = Agent\n",
    "\n",
    "# run experiment\n",
    "run_experiment(current_env, current_agent, environment_parameters, agent_parameters, experiment_parameters)"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "5e6d81cab5de679eba86a658acbe9237",
     "grade": false,
     "grade_id": "cell40",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Run the cell below to see the comparison between the agent that you implemented and a random agent for the one run and 300 episodes. Note that the `plot_result()` function smoothes the learning curve by applying a sliding window on the performance measure. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "image/svg+xml": "<?xml version=\"1.0\" encoding=\"utf-8\" standalone=\"no\"?>\n<!DOCTYPE svg PUBLIC \"-//W3C//DTD SVG 1.1//EN\"\n  \"http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd\">\n<!-- Created with matplotlib (https://matplotlib.org/) -->\n<svg height=\"423.425156pt\" version=\"1.1\" viewBox=\"0 0 568.95625 423.425156\" width=\"568.95625pt\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\">\n <defs>\n  <style type=\"text/css\">\n*{stroke-linecap:butt;stroke-linejoin:round;}\n  </style>\n </defs>\n <g id=\"figure_1\">\n  <g id=\"patch_1\">\n   <path d=\"M 0 423.425156 \nL 568.95625 423.425156 \nL 568.95625 0 \nL 0 0 \nz\n\" style=\"fill:none;\"/>\n  </g>\n  <g id=\"axes_1\">\n   <g id=\"patch_2\">\n    <path d=\"M 109.04375 381.997656 \nL 561.75625 381.997656 \nL 561.75625 24.597656 \nL 109.04375 24.597656 \nz\n\" style=\"fill:#ffffff;\"/>\n   </g>\n   <g id=\"matplotlib.axis_1\">\n    <g id=\"xtick_1\">\n     <g id=\"line2d_1\">\n      <defs>\n       <path d=\"M 0 0 \nL 0 3.5 \n\" id=\"m407d24431b\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n      </defs>\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"129.621591\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_1\">\n      <!-- 0 -->\n      <defs>\n       <path d=\"M 31.78125 66.40625 \nQ 24.171875 66.40625 20.328125 58.90625 \nQ 16.5 51.421875 16.5 36.375 \nQ 16.5 21.390625 20.328125 13.890625 \nQ 24.171875 6.390625 31.78125 6.390625 \nQ 39.453125 6.390625 43.28125 13.890625 \nQ 47.125 21.390625 47.125 36.375 \nQ 47.125 51.421875 43.28125 58.90625 \nQ 39.453125 66.40625 31.78125 66.40625 \nz\nM 31.78125 74.21875 \nQ 44.046875 74.21875 50.515625 64.515625 \nQ 56.984375 54.828125 56.984375 36.375 \nQ 56.984375 17.96875 50.515625 8.265625 \nQ 44.046875 -1.421875 31.78125 -1.421875 \nQ 19.53125 -1.421875 13.0625 8.265625 \nQ 6.59375 17.96875 6.59375 36.375 \nQ 6.59375 54.828125 13.0625 64.515625 \nQ 19.53125 74.21875 31.78125 74.21875 \nz\n\" id=\"DejaVuSans-48\"/>\n      </defs>\n      <g transform=\"translate(126.440341 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_2\">\n     <g id=\"line2d_2\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"198.443801\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_2\">\n      <!-- 50 -->\n      <defs>\n       <path d=\"M 10.796875 72.90625 \nL 49.515625 72.90625 \nL 49.515625 64.59375 \nL 19.828125 64.59375 \nL 19.828125 46.734375 \nQ 21.96875 47.46875 24.109375 47.828125 \nQ 26.265625 48.1875 28.421875 48.1875 \nQ 40.625 48.1875 47.75 41.5 \nQ 54.890625 34.8125 54.890625 23.390625 \nQ 54.890625 11.625 47.5625 5.09375 \nQ 40.234375 -1.421875 26.90625 -1.421875 \nQ 22.3125 -1.421875 17.546875 -0.640625 \nQ 12.796875 0.140625 7.71875 1.703125 \nL 7.71875 11.625 \nQ 12.109375 9.234375 16.796875 8.0625 \nQ 21.484375 6.890625 26.703125 6.890625 \nQ 35.15625 6.890625 40.078125 11.328125 \nQ 45.015625 15.765625 45.015625 23.390625 \nQ 45.015625 31 40.078125 35.4375 \nQ 35.15625 39.890625 26.703125 39.890625 \nQ 22.75 39.890625 18.8125 39.015625 \nQ 14.890625 38.140625 10.796875 36.28125 \nz\n\" id=\"DejaVuSans-53\"/>\n      </defs>\n      <g transform=\"translate(192.081301 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-53\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_3\">\n     <g id=\"line2d_3\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"267.266012\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_3\">\n      <!-- 100 -->\n      <defs>\n       <path d=\"M 12.40625 8.296875 \nL 28.515625 8.296875 \nL 28.515625 63.921875 \nL 10.984375 60.40625 \nL 10.984375 69.390625 \nL 28.421875 72.90625 \nL 38.28125 72.90625 \nL 38.28125 8.296875 \nL 54.390625 8.296875 \nL 54.390625 0 \nL 12.40625 0 \nz\n\" id=\"DejaVuSans-49\"/>\n      </defs>\n      <g transform=\"translate(257.722262 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-49\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_4\">\n     <g id=\"line2d_4\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"336.088222\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_4\">\n      <!-- 150 -->\n      <g transform=\"translate(326.544472 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-49\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-53\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_5\">\n     <g id=\"line2d_5\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"404.910433\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_5\">\n      <!-- 200 -->\n      <defs>\n       <path d=\"M 19.1875 8.296875 \nL 53.609375 8.296875 \nL 53.609375 0 \nL 7.328125 0 \nL 7.328125 8.296875 \nQ 12.9375 14.109375 22.625 23.890625 \nQ 32.328125 33.6875 34.8125 36.53125 \nQ 39.546875 41.84375 41.421875 45.53125 \nQ 43.3125 49.21875 43.3125 52.78125 \nQ 43.3125 58.59375 39.234375 62.25 \nQ 35.15625 65.921875 28.609375 65.921875 \nQ 23.96875 65.921875 18.8125 64.3125 \nQ 13.671875 62.703125 7.8125 59.421875 \nL 7.8125 69.390625 \nQ 13.765625 71.78125 18.9375 73 \nQ 24.125 74.21875 28.421875 74.21875 \nQ 39.75 74.21875 46.484375 68.546875 \nQ 53.21875 62.890625 53.21875 53.421875 \nQ 53.21875 48.921875 51.53125 44.890625 \nQ 49.859375 40.875 45.40625 35.40625 \nQ 44.1875 33.984375 37.640625 27.21875 \nQ 31.109375 20.453125 19.1875 8.296875 \nz\n\" id=\"DejaVuSans-50\"/>\n      </defs>\n      <g transform=\"translate(395.366683 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-50\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_6\">\n     <g id=\"line2d_6\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"473.732643\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_6\">\n      <!-- 250 -->\n      <g transform=\"translate(464.188893 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-50\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-53\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"xtick_7\">\n     <g id=\"line2d_7\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"542.554853\" xlink:href=\"#m407d24431b\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_7\">\n      <!-- 300 -->\n      <defs>\n       <path d=\"M 40.578125 39.3125 \nQ 47.65625 37.796875 51.625 33 \nQ 55.609375 28.21875 55.609375 21.1875 \nQ 55.609375 10.40625 48.1875 4.484375 \nQ 40.765625 -1.421875 27.09375 -1.421875 \nQ 22.515625 -1.421875 17.65625 -0.515625 \nQ 12.796875 0.390625 7.625 2.203125 \nL 7.625 11.71875 \nQ 11.71875 9.328125 16.59375 8.109375 \nQ 21.484375 6.890625 26.8125 6.890625 \nQ 36.078125 6.890625 40.9375 10.546875 \nQ 45.796875 14.203125 45.796875 21.1875 \nQ 45.796875 27.640625 41.28125 31.265625 \nQ 36.765625 34.90625 28.71875 34.90625 \nL 20.21875 34.90625 \nL 20.21875 43.015625 \nL 29.109375 43.015625 \nQ 36.375 43.015625 40.234375 45.921875 \nQ 44.09375 48.828125 44.09375 54.296875 \nQ 44.09375 59.90625 40.109375 62.90625 \nQ 36.140625 65.921875 28.71875 65.921875 \nQ 24.65625 65.921875 20.015625 65.03125 \nQ 15.375 64.15625 9.8125 62.3125 \nL 9.8125 71.09375 \nQ 15.4375 72.65625 20.34375 73.4375 \nQ 25.25 74.21875 29.59375 74.21875 \nQ 40.828125 74.21875 47.359375 69.109375 \nQ 53.90625 64.015625 53.90625 55.328125 \nQ 53.90625 49.265625 50.4375 45.09375 \nQ 46.96875 40.921875 40.578125 39.3125 \nz\n\" id=\"DejaVuSans-51\"/>\n      </defs>\n      <g transform=\"translate(533.011103 396.596094)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-51\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"text_8\">\n     <!-- Episodes -->\n     <defs>\n      <path d=\"M 9.8125 72.90625 \nL 55.90625 72.90625 \nL 55.90625 64.59375 \nL 19.671875 64.59375 \nL 19.671875 43.015625 \nL 54.390625 43.015625 \nL 54.390625 34.71875 \nL 19.671875 34.71875 \nL 19.671875 8.296875 \nL 56.78125 8.296875 \nL 56.78125 0 \nL 9.8125 0 \nz\n\" id=\"DejaVuSans-69\"/>\n      <path d=\"M 18.109375 8.203125 \nL 18.109375 -20.796875 \nL 9.078125 -20.796875 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.390625 \nQ 20.953125 51.265625 25.265625 53.625 \nQ 29.59375 56 35.59375 56 \nQ 45.5625 56 51.78125 48.09375 \nQ 58.015625 40.1875 58.015625 27.296875 \nQ 58.015625 14.40625 51.78125 6.484375 \nQ 45.5625 -1.421875 35.59375 -1.421875 \nQ 29.59375 -1.421875 25.265625 0.953125 \nQ 20.953125 3.328125 18.109375 8.203125 \nz\nM 48.6875 27.296875 \nQ 48.6875 37.203125 44.609375 42.84375 \nQ 40.53125 48.484375 33.40625 48.484375 \nQ 26.265625 48.484375 22.1875 42.84375 \nQ 18.109375 37.203125 18.109375 27.296875 \nQ 18.109375 17.390625 22.1875 11.75 \nQ 26.265625 6.109375 33.40625 6.109375 \nQ 40.53125 6.109375 44.609375 11.75 \nQ 48.6875 17.390625 48.6875 27.296875 \nz\n\" id=\"DejaVuSans-112\"/>\n      <path d=\"M 9.421875 54.6875 \nL 18.40625 54.6875 \nL 18.40625 0 \nL 9.421875 0 \nz\nM 9.421875 75.984375 \nL 18.40625 75.984375 \nL 18.40625 64.59375 \nL 9.421875 64.59375 \nz\n\" id=\"DejaVuSans-105\"/>\n      <path d=\"M 44.28125 53.078125 \nL 44.28125 44.578125 \nQ 40.484375 46.53125 36.375 47.5 \nQ 32.28125 48.484375 27.875 48.484375 \nQ 21.1875 48.484375 17.84375 46.4375 \nQ 14.5 44.390625 14.5 40.28125 \nQ 14.5 37.15625 16.890625 35.375 \nQ 19.28125 33.59375 26.515625 31.984375 \nL 29.59375 31.296875 \nQ 39.15625 29.25 43.1875 25.515625 \nQ 47.21875 21.78125 47.21875 15.09375 \nQ 47.21875 7.46875 41.1875 3.015625 \nQ 35.15625 -1.421875 24.609375 -1.421875 \nQ 20.21875 -1.421875 15.453125 -0.5625 \nQ 10.6875 0.296875 5.421875 2 \nL 5.421875 11.28125 \nQ 10.40625 8.6875 15.234375 7.390625 \nQ 20.0625 6.109375 24.8125 6.109375 \nQ 31.15625 6.109375 34.5625 8.28125 \nQ 37.984375 10.453125 37.984375 14.40625 \nQ 37.984375 18.0625 35.515625 20.015625 \nQ 33.0625 21.96875 24.703125 23.78125 \nL 21.578125 24.515625 \nQ 13.234375 26.265625 9.515625 29.90625 \nQ 5.8125 33.546875 5.8125 39.890625 \nQ 5.8125 47.609375 11.28125 51.796875 \nQ 16.75 56 26.8125 56 \nQ 31.78125 56 36.171875 55.265625 \nQ 40.578125 54.546875 44.28125 53.078125 \nz\n\" id=\"DejaVuSans-115\"/>\n      <path d=\"M 30.609375 48.390625 \nQ 23.390625 48.390625 19.1875 42.75 \nQ 14.984375 37.109375 14.984375 27.296875 \nQ 14.984375 17.484375 19.15625 11.84375 \nQ 23.34375 6.203125 30.609375 6.203125 \nQ 37.796875 6.203125 41.984375 11.859375 \nQ 46.1875 17.53125 46.1875 27.296875 \nQ 46.1875 37.015625 41.984375 42.703125 \nQ 37.796875 48.390625 30.609375 48.390625 \nz\nM 30.609375 56 \nQ 42.328125 56 49.015625 48.375 \nQ 55.71875 40.765625 55.71875 27.296875 \nQ 55.71875 13.875 49.015625 6.21875 \nQ 42.328125 -1.421875 30.609375 -1.421875 \nQ 18.84375 -1.421875 12.171875 6.21875 \nQ 5.515625 13.875 5.515625 27.296875 \nQ 5.515625 40.765625 12.171875 48.375 \nQ 18.84375 56 30.609375 56 \nz\n\" id=\"DejaVuSans-111\"/>\n      <path d=\"M 45.40625 46.390625 \nL 45.40625 75.984375 \nL 54.390625 75.984375 \nL 54.390625 0 \nL 45.40625 0 \nL 45.40625 8.203125 \nQ 42.578125 3.328125 38.25 0.953125 \nQ 33.9375 -1.421875 27.875 -1.421875 \nQ 17.96875 -1.421875 11.734375 6.484375 \nQ 5.515625 14.40625 5.515625 27.296875 \nQ 5.515625 40.1875 11.734375 48.09375 \nQ 17.96875 56 27.875 56 \nQ 33.9375 56 38.25 53.625 \nQ 42.578125 51.265625 45.40625 46.390625 \nz\nM 14.796875 27.296875 \nQ 14.796875 17.390625 18.875 11.75 \nQ 22.953125 6.109375 30.078125 6.109375 \nQ 37.203125 6.109375 41.296875 11.75 \nQ 45.40625 17.390625 45.40625 27.296875 \nQ 45.40625 37.203125 41.296875 42.84375 \nQ 37.203125 48.484375 30.078125 48.484375 \nQ 22.953125 48.484375 18.875 42.84375 \nQ 14.796875 37.203125 14.796875 27.296875 \nz\n\" id=\"DejaVuSans-100\"/>\n      <path d=\"M 56.203125 29.59375 \nL 56.203125 25.203125 \nL 14.890625 25.203125 \nQ 15.484375 15.921875 20.484375 11.0625 \nQ 25.484375 6.203125 34.421875 6.203125 \nQ 39.59375 6.203125 44.453125 7.46875 \nQ 49.3125 8.734375 54.109375 11.28125 \nL 54.109375 2.78125 \nQ 49.265625 0.734375 44.1875 -0.34375 \nQ 39.109375 -1.421875 33.890625 -1.421875 \nQ 20.796875 -1.421875 13.15625 6.1875 \nQ 5.515625 13.8125 5.515625 26.8125 \nQ 5.515625 40.234375 12.765625 48.109375 \nQ 20.015625 56 32.328125 56 \nQ 43.359375 56 49.78125 48.890625 \nQ 56.203125 41.796875 56.203125 29.59375 \nz\nM 47.21875 32.234375 \nQ 47.125 39.59375 43.09375 43.984375 \nQ 39.0625 48.390625 32.421875 48.390625 \nQ 24.90625 48.390625 20.390625 44.140625 \nQ 15.875 39.890625 15.1875 32.171875 \nz\n\" id=\"DejaVuSans-101\"/>\n     </defs>\n     <g transform=\"translate(304.260937 413.313594)scale(0.14 -0.14)\">\n      <use xlink:href=\"#DejaVuSans-69\"/>\n      <use x=\"63.183594\" xlink:href=\"#DejaVuSans-112\"/>\n      <use x=\"126.660156\" xlink:href=\"#DejaVuSans-105\"/>\n      <use x=\"154.443359\" xlink:href=\"#DejaVuSans-115\"/>\n      <use x=\"206.542969\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"267.724609\" xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"331.201172\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"392.724609\" xlink:href=\"#DejaVuSans-115\"/>\n     </g>\n    </g>\n   </g>\n   <g id=\"matplotlib.axis_2\">\n    <g id=\"ytick_1\">\n     <g id=\"line2d_8\">\n      <defs>\n       <path d=\"M 0 0 \nL -3.5 0 \n\" id=\"m02954099c2\" style=\"stroke:#000000;stroke-width:0.8;\"/>\n      </defs>\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"381.997656\"/>\n      </g>\n     </g>\n     <g id=\"text_9\">\n      <!-- −300 -->\n      <defs>\n       <path d=\"M 10.59375 35.5 \nL 73.1875 35.5 \nL 73.1875 27.203125 \nL 10.59375 27.203125 \nz\n\" id=\"DejaVuSans-8722\"/>\n      </defs>\n      <g transform=\"translate(74.576562 385.796875)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-8722\"/>\n       <use x=\"83.789062\" xlink:href=\"#DejaVuSans-51\"/>\n       <use x=\"147.412109\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"211.035156\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_2\">\n     <g id=\"line2d_9\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"322.43099\"/>\n      </g>\n     </g>\n     <g id=\"text_10\">\n      <!-- −200 -->\n      <g transform=\"translate(74.576562 326.230208)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-8722\"/>\n       <use x=\"83.789062\" xlink:href=\"#DejaVuSans-50\"/>\n       <use x=\"147.412109\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"211.035156\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_3\">\n     <g id=\"line2d_10\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"262.864323\"/>\n      </g>\n     </g>\n     <g id=\"text_11\">\n      <!-- −100 -->\n      <g transform=\"translate(74.576562 266.663542)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-8722\"/>\n       <use x=\"83.789062\" xlink:href=\"#DejaVuSans-49\"/>\n       <use x=\"147.412109\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"211.035156\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_4\">\n     <g id=\"line2d_11\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"203.297656\"/>\n      </g>\n     </g>\n     <g id=\"text_12\">\n      <!-- 0 -->\n      <g transform=\"translate(95.68125 207.096875)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_5\">\n     <g id=\"line2d_12\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"143.73099\"/>\n      </g>\n     </g>\n     <g id=\"text_13\">\n      <!-- 100 -->\n      <g transform=\"translate(82.95625 147.530208)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-49\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_6\">\n     <g id=\"line2d_13\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"84.164323\"/>\n      </g>\n     </g>\n     <g id=\"text_14\">\n      <!-- 200 -->\n      <g transform=\"translate(82.95625 87.963542)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-50\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"ytick_7\">\n     <g id=\"line2d_14\">\n      <g>\n       <use style=\"stroke:#000000;stroke-width:0.8;\" x=\"109.04375\" xlink:href=\"#m02954099c2\" y=\"24.597656\"/>\n      </g>\n     </g>\n     <g id=\"text_15\">\n      <!-- 300 -->\n      <g transform=\"translate(82.95625 28.396875)scale(0.1 -0.1)\">\n       <use xlink:href=\"#DejaVuSans-51\"/>\n       <use x=\"63.623047\" xlink:href=\"#DejaVuSans-48\"/>\n       <use x=\"127.246094\" xlink:href=\"#DejaVuSans-48\"/>\n      </g>\n     </g>\n    </g>\n    <g id=\"text_16\">\n     <!-- Sum of -->\n     <defs>\n      <path d=\"M 53.515625 70.515625 \nL 53.515625 60.890625 \nQ 47.90625 63.578125 42.921875 64.890625 \nQ 37.9375 66.21875 33.296875 66.21875 \nQ 25.25 66.21875 20.875 63.09375 \nQ 16.5 59.96875 16.5 54.203125 \nQ 16.5 49.359375 19.40625 46.890625 \nQ 22.3125 44.4375 30.421875 42.921875 \nL 36.375 41.703125 \nQ 47.40625 39.59375 52.65625 34.296875 \nQ 57.90625 29 57.90625 20.125 \nQ 57.90625 9.515625 50.796875 4.046875 \nQ 43.703125 -1.421875 29.984375 -1.421875 \nQ 24.8125 -1.421875 18.96875 -0.25 \nQ 13.140625 0.921875 6.890625 3.21875 \nL 6.890625 13.375 \nQ 12.890625 10.015625 18.65625 8.296875 \nQ 24.421875 6.59375 29.984375 6.59375 \nQ 38.421875 6.59375 43.015625 9.90625 \nQ 47.609375 13.234375 47.609375 19.390625 \nQ 47.609375 24.75 44.3125 27.78125 \nQ 41.015625 30.8125 33.5 32.328125 \nL 27.484375 33.5 \nQ 16.453125 35.6875 11.515625 40.375 \nQ 6.59375 45.0625 6.59375 53.421875 \nQ 6.59375 63.09375 13.40625 68.65625 \nQ 20.21875 74.21875 32.171875 74.21875 \nQ 37.3125 74.21875 42.625 73.28125 \nQ 47.953125 72.359375 53.515625 70.515625 \nz\n\" id=\"DejaVuSans-83\"/>\n      <path d=\"M 8.5 21.578125 \nL 8.5 54.6875 \nL 17.484375 54.6875 \nL 17.484375 21.921875 \nQ 17.484375 14.15625 20.5 10.265625 \nQ 23.53125 6.390625 29.59375 6.390625 \nQ 36.859375 6.390625 41.078125 11.03125 \nQ 45.3125 15.671875 45.3125 23.6875 \nL 45.3125 54.6875 \nL 54.296875 54.6875 \nL 54.296875 0 \nL 45.3125 0 \nL 45.3125 8.40625 \nQ 42.046875 3.421875 37.71875 1 \nQ 33.40625 -1.421875 27.6875 -1.421875 \nQ 18.265625 -1.421875 13.375 4.4375 \nQ 8.5 10.296875 8.5 21.578125 \nz\nM 31.109375 56 \nz\n\" id=\"DejaVuSans-117\"/>\n      <path d=\"M 52 44.1875 \nQ 55.375 50.25 60.0625 53.125 \nQ 64.75 56 71.09375 56 \nQ 79.640625 56 84.28125 50.015625 \nQ 88.921875 44.046875 88.921875 33.015625 \nL 88.921875 0 \nL 79.890625 0 \nL 79.890625 32.71875 \nQ 79.890625 40.578125 77.09375 44.375 \nQ 74.3125 48.1875 68.609375 48.1875 \nQ 61.625 48.1875 57.5625 43.546875 \nQ 53.515625 38.921875 53.515625 30.90625 \nL 53.515625 0 \nL 44.484375 0 \nL 44.484375 32.71875 \nQ 44.484375 40.625 41.703125 44.40625 \nQ 38.921875 48.1875 33.109375 48.1875 \nQ 26.21875 48.1875 22.15625 43.53125 \nQ 18.109375 38.875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 21.1875 51.21875 25.484375 53.609375 \nQ 29.78125 56 35.6875 56 \nQ 41.65625 56 45.828125 52.96875 \nQ 50 49.953125 52 44.1875 \nz\n\" id=\"DejaVuSans-109\"/>\n      <path id=\"DejaVuSans-32\"/>\n      <path d=\"M 37.109375 75.984375 \nL 37.109375 68.5 \nL 28.515625 68.5 \nQ 23.6875 68.5 21.796875 66.546875 \nQ 19.921875 64.59375 19.921875 59.515625 \nL 19.921875 54.6875 \nL 34.71875 54.6875 \nL 34.71875 47.703125 \nL 19.921875 47.703125 \nL 19.921875 0 \nL 10.890625 0 \nL 10.890625 47.703125 \nL 2.296875 47.703125 \nL 2.296875 54.6875 \nL 10.890625 54.6875 \nL 10.890625 58.5 \nQ 10.890625 67.625 15.140625 71.796875 \nQ 19.390625 75.984375 28.609375 75.984375 \nz\n\" id=\"DejaVuSans-102\"/>\n     </defs>\n     <g transform=\"translate(9.905937 153.355281)scale(0.14 -0.14)\">\n      <use xlink:href=\"#DejaVuSans-83\"/>\n      <use x=\"63.476562\" xlink:href=\"#DejaVuSans-117\"/>\n      <use x=\"126.855469\" xlink:href=\"#DejaVuSans-109\"/>\n      <use x=\"224.267578\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"256.054688\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"317.236328\" xlink:href=\"#DejaVuSans-102\"/>\n     </g>\n     <!-- reward -->\n     <defs>\n      <path d=\"M 41.109375 46.296875 \nQ 39.59375 47.171875 37.8125 47.578125 \nQ 36.03125 48 33.890625 48 \nQ 26.265625 48 22.1875 43.046875 \nQ 18.109375 38.09375 18.109375 28.8125 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 20.953125 51.171875 25.484375 53.578125 \nQ 30.03125 56 36.53125 56 \nQ 37.453125 56 38.578125 55.875 \nQ 39.703125 55.765625 41.0625 55.515625 \nz\n\" id=\"DejaVuSans-114\"/>\n      <path d=\"M 4.203125 54.6875 \nL 13.1875 54.6875 \nL 24.421875 12.015625 \nL 35.59375 54.6875 \nL 46.1875 54.6875 \nL 57.421875 12.015625 \nL 68.609375 54.6875 \nL 77.59375 54.6875 \nL 63.28125 0 \nL 52.6875 0 \nL 40.921875 44.828125 \nL 29.109375 0 \nL 18.5 0 \nz\n\" id=\"DejaVuSans-119\"/>\n      <path d=\"M 34.28125 27.484375 \nQ 23.390625 27.484375 19.1875 25 \nQ 14.984375 22.515625 14.984375 16.5 \nQ 14.984375 11.71875 18.140625 8.90625 \nQ 21.296875 6.109375 26.703125 6.109375 \nQ 34.1875 6.109375 38.703125 11.40625 \nQ 43.21875 16.703125 43.21875 25.484375 \nL 43.21875 27.484375 \nz\nM 52.203125 31.203125 \nL 52.203125 0 \nL 43.21875 0 \nL 43.21875 8.296875 \nQ 40.140625 3.328125 35.546875 0.953125 \nQ 30.953125 -1.421875 24.3125 -1.421875 \nQ 15.921875 -1.421875 10.953125 3.296875 \nQ 6 8.015625 6 15.921875 \nQ 6 25.140625 12.171875 29.828125 \nQ 18.359375 34.515625 30.609375 34.515625 \nL 43.21875 34.515625 \nL 43.21875 35.40625 \nQ 43.21875 41.609375 39.140625 45 \nQ 35.0625 48.390625 27.6875 48.390625 \nQ 23 48.390625 18.546875 47.265625 \nQ 14.109375 46.140625 10.015625 43.890625 \nL 10.015625 52.203125 \nQ 14.9375 54.109375 19.578125 55.046875 \nQ 24.21875 56 28.609375 56 \nQ 40.484375 56 46.34375 49.84375 \nQ 52.203125 43.703125 52.203125 31.203125 \nz\n\" id=\"DejaVuSans-97\"/>\n     </defs>\n     <g transform=\"translate(10.059062 169.032219)scale(0.14 -0.14)\">\n      <use xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"41.082031\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"102.605469\" xlink:href=\"#DejaVuSans-119\"/>\n      <use x=\"184.392578\" xlink:href=\"#DejaVuSans-97\"/>\n      <use x=\"245.671875\" xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"286.769531\" xlink:href=\"#DejaVuSans-100\"/>\n     </g>\n     <!-- during -->\n     <defs>\n      <path d=\"M 54.890625 33.015625 \nL 54.890625 0 \nL 45.90625 0 \nL 45.90625 32.71875 \nQ 45.90625 40.484375 42.875 44.328125 \nQ 39.84375 48.1875 33.796875 48.1875 \nQ 26.515625 48.1875 22.3125 43.546875 \nQ 18.109375 38.921875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 54.6875 \nL 18.109375 54.6875 \nL 18.109375 46.1875 \nQ 21.34375 51.125 25.703125 53.5625 \nQ 30.078125 56 35.796875 56 \nQ 45.21875 56 50.046875 50.171875 \nQ 54.890625 44.34375 54.890625 33.015625 \nz\n\" id=\"DejaVuSans-110\"/>\n      <path d=\"M 45.40625 27.984375 \nQ 45.40625 37.75 41.375 43.109375 \nQ 37.359375 48.484375 30.078125 48.484375 \nQ 22.859375 48.484375 18.828125 43.109375 \nQ 14.796875 37.75 14.796875 27.984375 \nQ 14.796875 18.265625 18.828125 12.890625 \nQ 22.859375 7.515625 30.078125 7.515625 \nQ 37.359375 7.515625 41.375 12.890625 \nQ 45.40625 18.265625 45.40625 27.984375 \nz\nM 54.390625 6.78125 \nQ 54.390625 -7.171875 48.1875 -13.984375 \nQ 42 -20.796875 29.203125 -20.796875 \nQ 24.46875 -20.796875 20.265625 -20.09375 \nQ 16.0625 -19.390625 12.109375 -17.921875 \nL 12.109375 -9.1875 \nQ 16.0625 -11.328125 19.921875 -12.34375 \nQ 23.78125 -13.375 27.78125 -13.375 \nQ 36.625 -13.375 41.015625 -8.765625 \nQ 45.40625 -4.15625 45.40625 5.171875 \nL 45.40625 9.625 \nQ 42.625 4.78125 38.28125 2.390625 \nQ 33.9375 0 27.875 0 \nQ 17.828125 0 11.671875 7.65625 \nQ 5.515625 15.328125 5.515625 27.984375 \nQ 5.515625 40.671875 11.671875 48.328125 \nQ 17.828125 56 27.875 56 \nQ 33.9375 56 38.28125 53.609375 \nQ 42.625 51.21875 45.40625 46.390625 \nL 45.40625 54.6875 \nL 54.390625 54.6875 \nz\n\" id=\"DejaVuSans-103\"/>\n     </defs>\n     <g transform=\"translate(11.993906 184.709156)scale(0.14 -0.14)\">\n      <use xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"63.476562\" xlink:href=\"#DejaVuSans-117\"/>\n      <use x=\"126.855469\" xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"167.96875\" xlink:href=\"#DejaVuSans-105\"/>\n      <use x=\"195.751953\" xlink:href=\"#DejaVuSans-110\"/>\n      <use x=\"259.130859\" xlink:href=\"#DejaVuSans-103\"/>\n     </g>\n     <!-- episode -->\n     <g transform=\"translate(7.2 200.386094)scale(0.14 -0.14)\">\n      <use xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"61.523438\" xlink:href=\"#DejaVuSans-112\"/>\n      <use x=\"125\" xlink:href=\"#DejaVuSans-105\"/>\n      <use x=\"152.783203\" xlink:href=\"#DejaVuSans-115\"/>\n      <use x=\"204.882812\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"266.064453\" xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"329.541016\" xlink:href=\"#DejaVuSans-101\"/>\n     </g>\n    </g>\n   </g>\n   <g id=\"line2d_15\">\n    <path clip-path=\"url(#p88ee8b0c1a)\" d=\"M 129.621591 333.330836 \nL 130.998035 338.169931 \nL 132.374479 334.147622 \nL 133.750924 308.805593 \nL 135.127368 292.532849 \nL 136.503812 283.083687 \nL 137.880256 275.219707 \nL 139.2567 270.755951 \nL 140.633145 282.290842 \nL 142.009589 276.682356 \nL 143.386033 267.959606 \nL 144.762477 263.954198 \nL 146.138921 265.186399 \nL 147.515366 265.359716 \nL 148.89181 260.375901 \nL 150.268254 247.419314 \nL 151.644698 249.268843 \nL 153.021142 246.319541 \nL 154.397587 253.98902 \nL 155.774031 255.630252 \nL 157.150475 256.690958 \nL 158.526919 257.481377 \nL 159.903363 258.405893 \nL 161.279808 260.108763 \nL 162.656252 261.036327 \nL 164.032696 263.005142 \nL 165.40914 262.811064 \nL 166.785585 264.103153 \nL 168.162029 263.312438 \nL 170.914917 262.346857 \nL 172.291361 261.078239 \nL 173.667806 264.68775 \nL 175.04425 263.643649 \nL 176.420694 261.636726 \nL 177.797138 261.584103 \nL 179.173582 261.093103 \nL 180.550027 262.241487 \nL 181.926471 262.287405 \nL 183.302915 263.281684 \nL 184.679359 263.950152 \nL 186.055803 262.518759 \nL 187.432248 263.819257 \nL 188.808692 262.771771 \nL 190.185136 263.42141 \nL 191.56158 265.274179 \nL 195.690913 266.876753 \nL 197.067357 267.002561 \nL 198.443801 266.948044 \nL 199.820246 266.227634 \nL 202.573134 266.3345 \nL 203.949578 266.17074 \nL 205.326022 265.545609 \nL 208.078911 264.928985 \nL 209.455355 264.126475 \nL 210.831799 264.010253 \nL 214.961132 262.505642 \nL 216.337576 261.834945 \nL 221.843353 261.507088 \nL 223.219797 260.61816 \nL 224.596241 260.02211 \nL 227.34913 259.224656 \nL 228.725574 260.530726 \nL 230.102018 260.982326 \nL 231.478462 260.449405 \nL 232.854907 260.410452 \nL 235.607795 259.81814 \nL 238.360683 259.105644 \nL 239.737128 258.960873 \nL 242.490016 257.98643 \nL 243.86646 257.692932 \nL 247.995793 256.249131 \nL 249.372237 256.472898 \nL 250.748681 255.27528 \nL 252.125125 255.409962 \nL 253.50157 255.190626 \nL 254.878014 255.15396 \nL 256.254458 254.675192 \nL 257.630902 254.975068 \nL 259.007346 254.631326 \nL 260.383791 254.913037 \nL 261.760235 255.329582 \nL 264.513123 254.440447 \nL 265.889567 254.233154 \nL 267.266012 254.156228 \nL 270.0189 251.951574 \nL 271.395344 251.208424 \nL 274.148233 251.090941 \nL 275.524677 250.820077 \nL 278.277565 250.672577 \nL 279.65401 249.162471 \nL 281.030454 249.147386 \nL 282.406898 249.343239 \nL 283.783342 249.146671 \nL 286.536231 248.42645 \nL 287.912675 248.839904 \nL 289.289119 250.599561 \nL 290.665563 250.08909 \nL 292.042007 250.131509 \nL 293.418452 248.148265 \nL 294.794896 247.408216 \nL 296.17134 246.853943 \nL 297.547784 246.44769 \nL 298.924228 244.823048 \nL 301.677117 243.545371 \nL 303.053561 242.854437 \nL 304.430005 242.626474 \nL 305.80645 241.898663 \nL 307.182894 241.652596 \nL 308.559338 242.207402 \nL 309.935782 242.215802 \nL 311.312226 242.042103 \nL 312.688671 240.485401 \nL 314.065115 240.463683 \nL 315.441559 240.85366 \nL 316.818003 240.342215 \nL 318.194447 240.078308 \nL 319.570892 239.216645 \nL 320.947336 243.244389 \nL 322.32378 242.437102 \nL 323.700224 241.811244 \nL 325.076668 243.376228 \nL 326.453113 243.344215 \nL 327.829557 241.616178 \nL 329.206001 240.9309 \nL 330.582445 239.724868 \nL 331.958889 238.678506 \nL 333.335334 236.270931 \nL 334.711778 234.221268 \nL 336.088222 233.779261 \nL 337.464666 233.132942 \nL 338.841111 233.148343 \nL 340.217555 232.774033 \nL 341.593999 231.549854 \nL 342.970443 231.123478 \nL 344.346887 231.070355 \nL 345.723332 230.69352 \nL 347.099776 229.7204 \nL 348.47622 229.519875 \nL 349.852664 227.510281 \nL 351.229108 227.055642 \nL 352.605553 226.817844 \nL 353.981997 226.440527 \nL 356.734885 227.15477 \nL 358.111329 228.046353 \nL 359.487774 229.176113 \nL 360.864218 229.890959 \nL 363.617106 232.893792 \nL 364.99355 233.535883 \nL 367.746439 230.325817 \nL 369.122883 230.720536 \nL 370.499327 230.726935 \nL 371.875772 230.393019 \nL 374.62866 231.675203 \nL 376.005104 231.300537 \nL 377.381548 230.136191 \nL 378.757993 229.681337 \nL 380.134437 230.328908 \nL 381.510881 230.459105 \nL 382.887325 230.202501 \nL 384.263769 231.54358 \nL 385.640214 231.817591 \nL 387.016658 231.879049 \nL 388.393102 231.494003 \nL 389.769546 232.013034 \nL 391.14599 231.592936 \nL 392.522435 231.389148 \nL 395.275323 231.246263 \nL 396.651767 231.785915 \nL 398.028211 231.661798 \nL 400.7811 230.986974 \nL 402.157544 231.315552 \nL 403.533988 231.880137 \nL 404.910433 230.302758 \nL 406.286877 229.808429 \nL 407.663321 228.278919 \nL 410.416209 227.805358 \nL 411.792654 227.179603 \nL 413.169098 225.797218 \nL 414.545542 225.751983 \nL 415.921986 225.513806 \nL 417.29843 227.054407 \nL 418.674875 227.058449 \nL 420.051319 228.550806 \nL 421.427763 227.518359 \nL 422.804207 226.331816 \nL 424.180651 224.272952 \nL 425.557096 224.050948 \nL 426.93354 224.517556 \nL 428.309984 224.526306 \nL 429.686428 225.012175 \nL 431.062872 225.254818 \nL 432.439317 223.8507 \nL 433.815761 223.956815 \nL 435.192205 223.797894 \nL 436.568649 222.049892 \nL 437.945093 221.612813 \nL 440.697982 221.082621 \nL 442.074426 220.710731 \nL 446.203759 215.06497 \nL 447.580203 212.621226 \nL 448.956647 210.599389 \nL 450.333091 209.327286 \nL 451.709536 209.46909 \nL 453.08598 209.498192 \nL 455.838868 209.996063 \nL 457.215312 209.619834 \nL 458.591757 209.652535 \nL 459.968201 205.000116 \nL 461.344645 203.980286 \nL 462.721089 202.226926 \nL 464.097533 199.153927 \nL 465.473978 196.512544 \nL 466.850422 196.575679 \nL 468.226866 194.883386 \nL 469.60331 192.966979 \nL 470.979754 191.873233 \nL 472.356199 192.008967 \nL 473.732643 193.200616 \nL 475.109087 191.78216 \nL 476.485531 191.708082 \nL 477.861976 190.345974 \nL 480.614864 189.42441 \nL 481.991308 187.784785 \nL 483.367752 185.941164 \nL 484.744197 184.32171 \nL 486.120641 183.655465 \nL 487.497085 182.289793 \nL 488.873529 183.096427 \nL 493.002862 178.779146 \nL 495.75575 174.61991 \nL 497.132194 172.107716 \nL 498.508639 169.023775 \nL 499.885083 167.624415 \nL 502.637971 161.691717 \nL 504.014415 159.666117 \nL 505.39086 159.228904 \nL 506.767304 157.618858 \nL 508.143748 154.689857 \nL 509.520192 152.983749 \nL 510.896637 153.259478 \nL 513.649525 147.902673 \nL 515.025969 146.70009 \nL 516.402413 146.065275 \nL 517.778858 144.532443 \nL 519.155302 143.551586 \nL 520.531746 141.590067 \nL 521.90819 139.983855 \nL 523.284634 136.960995 \nL 524.661079 134.876522 \nL 526.037523 134.892951 \nL 527.413967 132.781978 \nL 528.790411 131.332555 \nL 531.5433 128.083656 \nL 532.919744 127.082765 \nL 534.296188 125.704908 \nL 535.672632 122.805727 \nL 537.049076 121.11687 \nL 541.178409 115.033647 \nL 541.178409 115.033647 \n\" style=\"fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:1.5;\"/>\n   </g>\n   <g id=\"line2d_16\">\n    <path clip-path=\"url(#p88ee8b0c1a)\" d=\"M 129.621591 269.097071 \nL 130.998035 286.267263 \nL 132.374479 281.80929 \nL 133.750924 318.30423 \nL 135.127368 360.087144 \nL 136.503812 354.96218 \nL 137.880256 342.371815 \nL 139.2567 334.77564 \nL 140.633145 342.735549 \nL 142.009589 336.317147 \nL 143.386033 333.137836 \nL 144.762477 340.998368 \nL 146.138921 342.058801 \nL 147.515366 348.876108 \nL 148.89181 347.786868 \nL 150.268254 342.636398 \nL 151.644698 339.729056 \nL 153.021142 335.432397 \nL 154.397587 332.190646 \nL 157.150475 337.788866 \nL 158.526919 338.357607 \nL 159.903363 337.225787 \nL 161.279808 334.926352 \nL 162.656252 331.410291 \nL 164.032696 328.81453 \nL 165.40914 330.337553 \nL 166.785585 330.365491 \nL 168.162029 327.680848 \nL 169.538473 326.991469 \nL 170.914917 326.671001 \nL 172.291361 324.068119 \nL 173.667806 322.998818 \nL 175.04425 327.005037 \nL 176.420694 325.500427 \nL 177.797138 322.352993 \nL 179.173582 321.223097 \nL 180.550027 322.752127 \nL 181.926471 321.501854 \nL 183.302915 322.099959 \nL 184.679359 321.241796 \nL 186.055803 321.607258 \nL 187.432248 320.216632 \nL 188.808692 321.875418 \nL 190.185136 320.213655 \nL 191.56158 321.06878 \nL 192.938024 319.467226 \nL 194.314469 320.058092 \nL 195.690913 318.8498 \nL 197.067357 320.178121 \nL 198.443801 319.238005 \nL 199.820246 321.080822 \nL 202.573134 319.096569 \nL 203.949578 318.374102 \nL 205.326022 318.111767 \nL 208.078911 316.122787 \nL 209.455355 317.306075 \nL 210.831799 316.258028 \nL 213.584688 319.771711 \nL 214.961132 320.625104 \nL 216.337576 321.196402 \nL 220.466909 318.25385 \nL 221.843353 317.608161 \nL 223.219797 318.550612 \nL 224.596241 318.208094 \nL 225.972685 317.325562 \nL 227.34913 318.564161 \nL 228.725574 318.015817 \nL 230.102018 318.207936 \nL 231.478462 318.058313 \nL 232.854907 319.125233 \nL 234.231351 319.189503 \nL 235.607795 318.572882 \nL 236.984239 318.940899 \nL 238.360683 319.505781 \nL 239.737128 320.903815 \nL 241.113572 321.034046 \nL 242.490016 321.703253 \nL 243.86646 321.003955 \nL 245.242904 321.960974 \nL 246.619349 322.197191 \nL 247.995793 321.855024 \nL 249.372237 320.949467 \nL 253.50157 319.35353 \nL 254.878014 318.888598 \nL 256.254458 318.145051 \nL 257.630902 317.714821 \nL 259.007346 318.499075 \nL 264.513123 316.047813 \nL 265.889567 315.266374 \nL 268.642456 318.454691 \nL 270.0189 318.837364 \nL 271.395344 318.764701 \nL 272.771789 317.401175 \nL 274.148233 316.515853 \nL 275.524677 317.470742 \nL 276.901121 316.965914 \nL 278.277565 317.992159 \nL 279.65401 316.800758 \nL 281.030454 318.042437 \nL 282.406898 318.176218 \nL 283.783342 317.331787 \nL 285.159786 316.750821 \nL 286.536231 314.919733 \nL 287.912675 315.370027 \nL 289.289119 316.473469 \nL 290.665563 316.103385 \nL 292.042007 316.274737 \nL 293.418452 316.1504 \nL 294.794896 316.657974 \nL 296.17134 315.182391 \nL 297.547784 314.085631 \nL 298.924228 313.817181 \nL 300.300673 314.34967 \nL 301.677117 315.453654 \nL 303.053561 314.759923 \nL 304.430005 314.783245 \nL 305.80645 313.935674 \nL 307.182894 314.253463 \nL 308.559338 314.271008 \nL 309.935782 314.448719 \nL 311.312226 315.859977 \nL 312.688671 315.358773 \nL 314.065115 314.007102 \nL 315.441559 314.600233 \nL 316.818003 315.048289 \nL 318.194447 315.320203 \nL 319.570892 314.28435 \nL 320.947336 315.509323 \nL 322.32378 315.110711 \nL 323.700224 314.8985 \nL 325.076668 314.256876 \nL 326.453113 314.302306 \nL 327.829557 313.277954 \nL 329.206001 313.294212 \nL 331.958889 316.022387 \nL 333.335334 315.190142 \nL 334.711778 316.434224 \nL 336.088222 315.954996 \nL 337.464666 316.007743 \nL 341.593999 315.782192 \nL 342.970443 316.392334 \nL 344.346887 315.200474 \nL 345.723332 315.54151 \nL 347.099776 315.213876 \nL 348.47622 315.78115 \nL 349.852664 317.59933 \nL 352.605553 314.466657 \nL 353.981997 313.405709 \nL 356.734885 313.400301 \nL 358.111329 313.656202 \nL 359.487774 313.470948 \nL 360.864218 313.461287 \nL 362.240662 313.057441 \nL 363.617106 313.993505 \nL 364.99355 315.805665 \nL 366.369995 314.645373 \nL 367.746439 315.08334 \nL 369.122883 314.546664 \nL 370.499327 314.446951 \nL 371.875772 313.354967 \nL 373.252216 312.58397 \nL 374.62866 312.352162 \nL 376.005104 311.282305 \nL 377.381548 311.40853 \nL 378.757993 310.006234 \nL 381.510881 308.185666 \nL 382.887325 309.689025 \nL 384.263769 309.830696 \nL 387.016658 309.722645 \nL 388.393102 311.201873 \nL 389.769546 311.4934 \nL 391.14599 311.942725 \nL 392.522435 313.019288 \nL 393.898879 312.856786 \nL 395.275323 313.789858 \nL 396.651767 313.63857 \nL 398.028211 314.240439 \nL 399.404656 314.117638 \nL 400.7811 314.662878 \nL 402.157544 316.107622 \nL 403.533988 318.616756 \nL 404.910433 319.228695 \nL 406.286877 317.245596 \nL 407.663321 315.593855 \nL 409.039765 315.240647 \nL 410.416209 315.09121 \nL 411.792654 314.206108 \nL 413.169098 312.093536 \nL 414.545542 310.776826 \nL 415.921986 311.542377 \nL 417.29843 312.097102 \nL 418.674875 311.786588 \nL 421.427763 309.289012 \nL 422.804207 308.255065 \nL 424.180651 307.92876 \nL 425.557096 307.892413 \nL 426.93354 306.977795 \nL 428.309984 305.828636 \nL 429.686428 306.12589 \nL 431.062872 305.327698 \nL 432.439317 306.927119 \nL 433.815761 305.259009 \nL 435.192205 306.248837 \nL 437.945093 306.694923 \nL 439.321538 305.913407 \nL 440.697982 304.93646 \nL 442.074426 307.06729 \nL 443.45087 306.348936 \nL 444.827315 306.371352 \nL 446.203759 307.484524 \nL 450.333091 306.101763 \nL 451.709536 306.150507 \nL 453.08598 306.610745 \nL 454.462424 305.43577 \nL 455.838868 306.237648 \nL 457.215312 305.682616 \nL 458.591757 306.110145 \nL 459.968201 305.192237 \nL 461.344645 304.108291 \nL 462.721089 305.214214 \nL 464.097533 305.182229 \nL 465.473978 305.854154 \nL 466.850422 305.953618 \nL 468.226866 306.932122 \nL 470.979754 304.431174 \nL 472.356199 304.511107 \nL 473.732643 303.562546 \nL 475.109087 303.850764 \nL 476.485531 304.254077 \nL 479.23842 304.44982 \nL 480.614864 304.630773 \nL 481.991308 303.435927 \nL 483.367752 304.183421 \nL 484.744197 305.244218 \nL 486.120641 304.928222 \nL 487.497085 304.332317 \nL 488.873529 303.310447 \nL 490.249973 303.305718 \nL 491.626418 303.760336 \nL 493.002862 304.741981 \nL 494.379306 303.994931 \nL 495.75575 304.477922 \nL 497.132194 304.22915 \nL 498.508639 304.842351 \nL 501.261527 304.967887 \nL 504.014415 302.161085 \nL 505.39086 303.319193 \nL 506.767304 303.206371 \nL 508.143748 303.470333 \nL 510.896637 302.807372 \nL 512.273081 302.94554 \nL 513.649525 303.973036 \nL 515.025969 304.142344 \nL 516.402413 303.148341 \nL 517.778858 302.874616 \nL 519.155302 303.028323 \nL 520.531746 302.879405 \nL 521.90819 302.011132 \nL 523.284634 299.937886 \nL 524.661079 301.045281 \nL 526.037523 300.749285 \nL 527.413967 301.255826 \nL 528.790411 301.52157 \nL 530.166855 301.415417 \nL 531.5433 300.331898 \nL 532.919744 300.22332 \nL 534.296188 299.42881 \nL 535.672632 299.260012 \nL 537.049076 297.538353 \nL 538.425521 297.766711 \nL 539.801965 297.254215 \nL 541.178409 296.394891 \nL 541.178409 296.394891 \n\" style=\"fill:none;stroke:#ff7f0e;stroke-linecap:square;stroke-width:1.5;\"/>\n   </g>\n   <g id=\"patch_3\">\n    <path d=\"M 109.04375 381.997656 \nL 109.04375 24.597656 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_4\">\n    <path d=\"M 561.75625 381.997656 \nL 561.75625 24.597656 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_5\">\n    <path d=\"M 109.04375 381.997656 \nL 561.75625 381.997656 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"patch_6\">\n    <path d=\"M 109.04375 24.597656 \nL 561.75625 24.597656 \n\" style=\"fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;\"/>\n   </g>\n   <g id=\"text_17\">\n    <!-- Learning Curve -->\n    <defs>\n     <path d=\"M 9.8125 72.90625 \nL 19.671875 72.90625 \nL 19.671875 8.296875 \nL 55.171875 8.296875 \nL 55.171875 0 \nL 9.8125 0 \nz\n\" id=\"DejaVuSans-76\"/>\n     <path d=\"M 64.40625 67.28125 \nL 64.40625 56.890625 \nQ 59.421875 61.53125 53.78125 63.8125 \nQ 48.140625 66.109375 41.796875 66.109375 \nQ 29.296875 66.109375 22.65625 58.46875 \nQ 16.015625 50.828125 16.015625 36.375 \nQ 16.015625 21.96875 22.65625 14.328125 \nQ 29.296875 6.6875 41.796875 6.6875 \nQ 48.140625 6.6875 53.78125 8.984375 \nQ 59.421875 11.28125 64.40625 15.921875 \nL 64.40625 5.609375 \nQ 59.234375 2.09375 53.4375 0.328125 \nQ 47.65625 -1.421875 41.21875 -1.421875 \nQ 24.65625 -1.421875 15.125 8.703125 \nQ 5.609375 18.84375 5.609375 36.375 \nQ 5.609375 53.953125 15.125 64.078125 \nQ 24.65625 74.21875 41.21875 74.21875 \nQ 47.75 74.21875 53.53125 72.484375 \nQ 59.328125 70.75 64.40625 67.28125 \nz\n\" id=\"DejaVuSans-67\"/>\n     <path d=\"M 2.984375 54.6875 \nL 12.5 54.6875 \nL 29.59375 8.796875 \nL 46.6875 54.6875 \nL 56.203125 54.6875 \nL 35.6875 0 \nL 23.484375 0 \nz\n\" id=\"DejaVuSans-118\"/>\n    </defs>\n    <g transform=\"translate(278.067187 18.597656)scale(0.15 -0.15)\">\n     <use xlink:href=\"#DejaVuSans-76\"/>\n     <use x=\"55.697266\" xlink:href=\"#DejaVuSans-101\"/>\n     <use x=\"117.220703\" xlink:href=\"#DejaVuSans-97\"/>\n     <use x=\"178.5\" xlink:href=\"#DejaVuSans-114\"/>\n     <use x=\"219.597656\" xlink:href=\"#DejaVuSans-110\"/>\n     <use x=\"282.976562\" xlink:href=\"#DejaVuSans-105\"/>\n     <use x=\"310.759766\" xlink:href=\"#DejaVuSans-110\"/>\n     <use x=\"374.138672\" xlink:href=\"#DejaVuSans-103\"/>\n     <use x=\"437.615234\" xlink:href=\"#DejaVuSans-32\"/>\n     <use x=\"469.402344\" xlink:href=\"#DejaVuSans-67\"/>\n     <use x=\"539.226562\" xlink:href=\"#DejaVuSans-117\"/>\n     <use x=\"602.605469\" xlink:href=\"#DejaVuSans-114\"/>\n     <use x=\"643.71875\" xlink:href=\"#DejaVuSans-118\"/>\n     <use x=\"702.898438\" xlink:href=\"#DejaVuSans-101\"/>\n    </g>\n   </g>\n   <g id=\"legend_1\">\n    <g id=\"patch_7\">\n     <path d=\"M 268.774844 73.160781 \nL 552.65625 73.160781 \nQ 555.25625 73.160781 555.25625 70.560781 \nL 555.25625 33.697656 \nQ 555.25625 31.097656 552.65625 31.097656 \nL 268.774844 31.097656 \nQ 266.174844 31.097656 266.174844 33.697656 \nL 266.174844 70.560781 \nQ 266.174844 73.160781 268.774844 73.160781 \nz\n\" style=\"fill:#ffffff;opacity:0.8;stroke:#cccccc;stroke-linejoin:miter;\"/>\n    </g>\n    <g id=\"line2d_17\">\n     <path d=\"M 271.374844 41.625625 \nL 297.374844 41.625625 \n\" style=\"fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:1.5;\"/>\n    </g>\n    <g id=\"line2d_18\"/>\n    <g id=\"text_18\">\n     <!-- Expected SARSA with neural network -->\n     <defs>\n      <path d=\"M 54.890625 54.6875 \nL 35.109375 28.078125 \nL 55.90625 0 \nL 45.3125 0 \nL 29.390625 21.484375 \nL 13.484375 0 \nL 2.875 0 \nL 24.125 28.609375 \nL 4.6875 54.6875 \nL 15.28125 54.6875 \nL 29.78125 35.203125 \nL 44.28125 54.6875 \nz\n\" id=\"DejaVuSans-120\"/>\n      <path d=\"M 48.78125 52.59375 \nL 48.78125 44.1875 \nQ 44.96875 46.296875 41.140625 47.34375 \nQ 37.3125 48.390625 33.40625 48.390625 \nQ 24.65625 48.390625 19.8125 42.84375 \nQ 14.984375 37.3125 14.984375 27.296875 \nQ 14.984375 17.28125 19.8125 11.734375 \nQ 24.65625 6.203125 33.40625 6.203125 \nQ 37.3125 6.203125 41.140625 7.25 \nQ 44.96875 8.296875 48.78125 10.40625 \nL 48.78125 2.09375 \nQ 45.015625 0.34375 40.984375 -0.53125 \nQ 36.96875 -1.421875 32.421875 -1.421875 \nQ 20.0625 -1.421875 12.78125 6.34375 \nQ 5.515625 14.109375 5.515625 27.296875 \nQ 5.515625 40.671875 12.859375 48.328125 \nQ 20.21875 56 33.015625 56 \nQ 37.15625 56 41.109375 55.140625 \nQ 45.0625 54.296875 48.78125 52.59375 \nz\n\" id=\"DejaVuSans-99\"/>\n      <path d=\"M 18.3125 70.21875 \nL 18.3125 54.6875 \nL 36.8125 54.6875 \nL 36.8125 47.703125 \nL 18.3125 47.703125 \nL 18.3125 18.015625 \nQ 18.3125 11.328125 20.140625 9.421875 \nQ 21.96875 7.515625 27.59375 7.515625 \nL 36.8125 7.515625 \nL 36.8125 0 \nL 27.59375 0 \nQ 17.1875 0 13.234375 3.875 \nQ 9.28125 7.765625 9.28125 18.015625 \nL 9.28125 47.703125 \nL 2.6875 47.703125 \nL 2.6875 54.6875 \nL 9.28125 54.6875 \nL 9.28125 70.21875 \nz\n\" id=\"DejaVuSans-116\"/>\n      <path d=\"M 34.1875 63.1875 \nL 20.796875 26.90625 \nL 47.609375 26.90625 \nz\nM 28.609375 72.90625 \nL 39.796875 72.90625 \nL 67.578125 0 \nL 57.328125 0 \nL 50.6875 18.703125 \nL 17.828125 18.703125 \nL 11.1875 0 \nL 0.78125 0 \nz\n\" id=\"DejaVuSans-65\"/>\n      <path d=\"M 44.390625 34.1875 \nQ 47.5625 33.109375 50.5625 29.59375 \nQ 53.5625 26.078125 56.59375 19.921875 \nL 66.609375 0 \nL 56 0 \nL 46.6875 18.703125 \nQ 43.0625 26.03125 39.671875 28.421875 \nQ 36.28125 30.8125 30.421875 30.8125 \nL 19.671875 30.8125 \nL 19.671875 0 \nL 9.8125 0 \nL 9.8125 72.90625 \nL 32.078125 72.90625 \nQ 44.578125 72.90625 50.734375 67.671875 \nQ 56.890625 62.453125 56.890625 51.90625 \nQ 56.890625 45.015625 53.6875 40.46875 \nQ 50.484375 35.9375 44.390625 34.1875 \nz\nM 19.671875 64.796875 \nL 19.671875 38.921875 \nL 32.078125 38.921875 \nQ 39.203125 38.921875 42.84375 42.21875 \nQ 46.484375 45.515625 46.484375 51.90625 \nQ 46.484375 58.296875 42.84375 61.546875 \nQ 39.203125 64.796875 32.078125 64.796875 \nz\n\" id=\"DejaVuSans-82\"/>\n      <path d=\"M 54.890625 33.015625 \nL 54.890625 0 \nL 45.90625 0 \nL 45.90625 32.71875 \nQ 45.90625 40.484375 42.875 44.328125 \nQ 39.84375 48.1875 33.796875 48.1875 \nQ 26.515625 48.1875 22.3125 43.546875 \nQ 18.109375 38.921875 18.109375 30.90625 \nL 18.109375 0 \nL 9.078125 0 \nL 9.078125 75.984375 \nL 18.109375 75.984375 \nL 18.109375 46.1875 \nQ 21.34375 51.125 25.703125 53.5625 \nQ 30.078125 56 35.796875 56 \nQ 45.21875 56 50.046875 50.171875 \nQ 54.890625 44.34375 54.890625 33.015625 \nz\n\" id=\"DejaVuSans-104\"/>\n      <path d=\"M 9.421875 75.984375 \nL 18.40625 75.984375 \nL 18.40625 0 \nL 9.421875 0 \nz\n\" id=\"DejaVuSans-108\"/>\n      <path d=\"M 9.078125 75.984375 \nL 18.109375 75.984375 \nL 18.109375 31.109375 \nL 44.921875 54.6875 \nL 56.390625 54.6875 \nL 27.390625 29.109375 \nL 57.625 0 \nL 45.90625 0 \nL 18.109375 26.703125 \nL 18.109375 0 \nL 9.078125 0 \nz\n\" id=\"DejaVuSans-107\"/>\n     </defs>\n     <g transform=\"translate(307.774844 46.175625)scale(0.13 -0.13)\">\n      <use xlink:href=\"#DejaVuSans-69\"/>\n      <use x=\"63.183594\" xlink:href=\"#DejaVuSans-120\"/>\n      <use x=\"122.363281\" xlink:href=\"#DejaVuSans-112\"/>\n      <use x=\"185.839844\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"247.363281\" xlink:href=\"#DejaVuSans-99\"/>\n      <use x=\"302.34375\" xlink:href=\"#DejaVuSans-116\"/>\n      <use x=\"341.552734\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"403.076172\" xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"466.552734\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"498.339844\" xlink:href=\"#DejaVuSans-83\"/>\n      <use x=\"561.832031\" xlink:href=\"#DejaVuSans-65\"/>\n      <use x=\"630.240234\" xlink:href=\"#DejaVuSans-82\"/>\n      <use x=\"699.722656\" xlink:href=\"#DejaVuSans-83\"/>\n      <use x=\"763.214844\" xlink:href=\"#DejaVuSans-65\"/>\n      <use x=\"831.623047\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"863.410156\" xlink:href=\"#DejaVuSans-119\"/>\n      <use x=\"945.197266\" xlink:href=\"#DejaVuSans-105\"/>\n      <use x=\"972.980469\" xlink:href=\"#DejaVuSans-116\"/>\n      <use x=\"1012.189453\" xlink:href=\"#DejaVuSans-104\"/>\n      <use x=\"1075.568359\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"1107.355469\" xlink:href=\"#DejaVuSans-110\"/>\n      <use x=\"1170.734375\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"1232.257812\" xlink:href=\"#DejaVuSans-117\"/>\n      <use x=\"1295.636719\" xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"1336.75\" xlink:href=\"#DejaVuSans-97\"/>\n      <use x=\"1398.029297\" xlink:href=\"#DejaVuSans-108\"/>\n      <use x=\"1425.8125\" xlink:href=\"#DejaVuSans-32\"/>\n      <use x=\"1457.599609\" xlink:href=\"#DejaVuSans-110\"/>\n      <use x=\"1520.978516\" xlink:href=\"#DejaVuSans-101\"/>\n      <use x=\"1582.501953\" xlink:href=\"#DejaVuSans-116\"/>\n      <use x=\"1621.710938\" xlink:href=\"#DejaVuSans-119\"/>\n      <use x=\"1703.498047\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"1764.679688\" xlink:href=\"#DejaVuSans-114\"/>\n      <use x=\"1805.792969\" xlink:href=\"#DejaVuSans-107\"/>\n     </g>\n    </g>\n    <g id=\"line2d_19\">\n     <path d=\"M 271.374844 60.707188 \nL 297.374844 60.707188 \n\" style=\"fill:none;stroke:#ff7f0e;stroke-linecap:square;stroke-width:1.5;\"/>\n    </g>\n    <g id=\"line2d_20\"/>\n    <g id=\"text_19\">\n     <!-- Random -->\n     <g transform=\"translate(307.774844 65.257188)scale(0.13 -0.13)\">\n      <use xlink:href=\"#DejaVuSans-82\"/>\n      <use x=\"69.451172\" xlink:href=\"#DejaVuSans-97\"/>\n      <use x=\"130.730469\" xlink:href=\"#DejaVuSans-110\"/>\n      <use x=\"194.109375\" xlink:href=\"#DejaVuSans-100\"/>\n      <use x=\"257.585938\" xlink:href=\"#DejaVuSans-111\"/>\n      <use x=\"318.767578\" xlink:href=\"#DejaVuSans-109\"/>\n     </g>\n    </g>\n   </g>\n  </g>\n </g>\n <defs>\n  <clipPath id=\"p88ee8b0c1a\">\n   <rect height=\"357.4\" width=\"452.7125\" x=\"109.04375\" y=\"24.597656\"/>\n  </clipPath>\n </defs>\n</svg>\n",
      "text/plain": "<Figure size 576x432 with 1 Axes>"
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plot_result([\"expected_sarsa_agent\", \"random_agent\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "206b2c64f0e560a6dbfc71a9f578a2fb",
     "grade": false,
     "grade_id": "cell42",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "In the following cell you can visualize the performance of the agent with a correct implementation. As you can see, the agent initially crashes quite quickly (Episode 0). Then, the agent learns to avoid crashing by expending fuel and staying far above the ground. Finally however, it learns to land smoothly within the landing zone demarcated by the two flags (Episode 275)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": "<div align=\"middle\">\n<video width=\"80%\" controls>\n      <source src=\"ImplementYourAgent.mp4\" type=\"video/mp4\">\n</video></div>\n",
      "text/plain": "<IPython.core.display.HTML object>"
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "%%HTML\n",
    "<div align=\"middle\">\n",
    "<video width=\"80%\" controls>\n",
    "      <source src=\"ImplementYourAgent.mp4\" type=\"video/mp4\">\n",
    "</video></div>"
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "8f628bc12879a6ba295fc81d943861df",
     "grade": false,
     "grade_id": "cell44",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "In the learning curve above, you can see that sum of reward over episode has quite a high-variance at the beginning. However, the performance seems to be improving. The experiment that you ran was for 300 episodes and 1 run. To understand how the agent performs in the long run, we provide below the learning curve for the agent trained for 3000 episodes with performance averaged over 30 runs.\n",
    "<img src=\"3000_episodes.png\" alt=\"Drawing\" style=\"width: 500px;\"/>\n",
    "You can see that the agent learns a reasonably good policy within 3000 episodes, gaining sum of reward bigger than 200. Note that because of the high-variance in the agent performance, we also smoothed the learning curve. "
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "f9aa77980600f6eaa6b04bdf163b726f",
     "grade": false,
     "grade_id": "cell45",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "### Wrapping up! \n",
    "\n",
    "You have successfully implemented Course 4 Programming Assignment 2.\n",
    "\n",
    "You have implemented an **Expected Sarsa agent with a neural network and the Adam optimizer** and used it for solving the Lunar Lander problem! You implemented different components of the agent including:\n",
    "\n",
    "- a neural network for function approximation,\n",
    "- the Adam algorithm for optimizing the weights of the neural network,\n",
    "- a Softmax policy,\n",
    "- the replay steps for updating the action-value function using the experiences sampled from a replay buffer\n",
    "\n",
    "You tested the agent for a single parameter setting. In the next assignment, you will perform a parameter study on the step-size parameter to gain insight about the effect of step-size on the performance of your agent."
   ]
  },
  {
   "cell_type": "markdown",
   "execution_count": null,
   "metadata": {
    "deletable": false,
    "editable": false,
    "nbgrader": {
     "checksum": "fefec5bec50493fb9b92455d57e06f3e",
     "grade": false,
     "grade_id": "cell46",
     "locked": true,
     "schema_version": 1,
     "solution": false
    }
   },
   "outputs": [],
   "source": [
    "Note: Apart from using the `Submit` button in the notebook, you have to submit an additional zip file containing the 'npy' files that were generated from running the experiment cells. In order to do so:\n",
    "1. Generate the zip file by running the experiment cells in the notebook. On the top of the notebook, navigate to `File->Open` to open the directory view of this assignment. Select the checkbox next to `results.zip` and click on `Download.` Alternatively, you can download the results folder and run `zip -jr results.zip results/` (The flag 'j' is required by the grader!).\n",
    "2. Go to the \"My submission\" tab on the programming assignment and click on \"+ Create submission\".\n",
    "3. Click on \"PA2 Data-file Grader\" and upload your results.zip.\n",
    "\n",
    "***These account for 25% of the marks, so don't forget to do so!***"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "coursera": {
   "course_slug": "complete-reinforcement-learning-system",
   "graded_item_id": "8dMlx",
   "launcher_item_id": "4O5gG"
  },
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.6-final"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}