Overview

In this repository some of the most popular reinforcement learning approaches are implemented. They all follow the structure of one abstract agent, in such a way, that helper functions to compare agents can be used on all instances. These are meant to serve as a learning and comparison tool.

Table of Contents

• Agents
• Environments
• Experiments
• Helper Functions in: utils and train

How to use it

Examples of usage can be found in /experiments as well as an explanation for those.

Creating an environment with gym.make(<gymid>) or by importing environment from /environnments where all the working envs are imported.

from environments import FrozenLakeEnv
env = FrozenLakeEnv(map_name='8x8')

To test some reinforcement learning methods, you can either implement your own method and comply with the abstract_agent from rl_methods. With their class implementation it is also described, for what type of environment they should work.

from agents import SarsaAgent, QLearningAgent
sarsa_agent = SarsaAgent(env)
q_learning_agent = QLearningAgent(env)

To train those agents and to visualise the statistics that emerged during training, functions from /train and /utils are used.

from train import train_agents
from utils import visualize_training_results_for_agents

stats = train_agents(env, [sarsa_agent, q_learning_agent])
visualize_training_results_for_agents(stats)

Resources

• Reinforcement Learning - An Introduction (2nd Edition)
• Playing Atari with Deep Reinforcement Learning
• Curiosity-driven Exploration by Self-supervised Prediction
• Mastering the game of Go with deep neural networks and tree search
• Continuous control with deep reinforcement learning
• colab research DDPG