Skip to content

Latest commit

 

History

History
73 lines (66 loc) · 5.06 KB

README.md

File metadata and controls

73 lines (66 loc) · 5.06 KB
Raw

Reinforcement Learning Battle City

The following is a Python code that consists of a RL tank agent capable of playing Battle City arcade game. The agent has been trained using PPO RL implementation in Stable Baselines 3 and Open AI's Gymnasium API. The Battle City game code was adapted from Raitis' repository https://github.com/raitisg/battle-city-tanks. Additionally, to train the RL agent, an agent based on the A* algorithm was adapted and improved from Wong Mun Hou's repository https://github.com/munhouiani/AI-Battle-City/tree/master. Files are included so that the agent can be furtherly trained. The model that has been trained (download weights from https://drive.google.com/drive/folders/1DgOlP0Pr-otEAJC12qajCUuHw3icMhHb) is able to successfully pass some levels with good chances but there is still margin for improvement. Feel free to try your own strategies to further improve the agent's performance! 🚀

Getting Started

Cloning the Repository

To clone the repository and start using the RL Battle City, follow these steps:

  1. Open a terminal or command prompt
  2. Navigate to the directory where you want to clone the repository.
  3. Run the following command:
git clone https://github.com/danisotelo/RL_battle_city.git

Installing Dependencies

Before running the program, you need to install Anaconda (https://www.anaconda.com/download). Once you have installed it, open Anaconda Navigator and go to Environments > Create. You need to create a new environment (call it tanks for example), and then, select the Python package 3.10.13 (this is important because it might not work for more recent Python versions). Once you have activated the environment, go to Home > VS Code.

Now you must open a VS Code terminal and install the required dependencies:

pip install gymnasium
pip install stable-baselines3[extra]

Next, you need to install the gym_tanks custom package. In order to do this you should select the folder where you have cloned the repository and try running from the terminal:

pip install .

If errors arrise, then you need to install Visual Studio Community 2022 because some C++ specific files are required. Once you install it, access to the Visual Studio Installer, go to Modify and install the following workloads:

  • Desktop development with C++
  • Universal Windows Platform development
  • Python development
  • Linux Development with C++
  • Visual Studio extension development
  • Game Development with C++

Additionally, you should also install the following individual components:

  • C++ Build Insights
  • Windows 10/11 SDK
  • MSVC v142 Build tools - VS 2019 C++ x64/x86 (v14.29-16.11)
  • MSVC Libraries with Spectre Mitigations - VS 2019 C++ for x64/x86 (v14.29-16.11)

Try again running pip install ., now it should run without errors and the installation is finished!.

Optional CUDA Installation for GPU Use

If you want to train the agent using your GPU instead of your CPU, first run nvidia-smi in your terminal to check the CUDA version of your GPU (the following steps are suitable if your CUDA Version >= 11.8, if not, you should look for a different CUDA Toolkit and PyTorch version). Download CUDA Toolkit 11.8 from https://developer.nvidia.com/cuda-11-8-0-download-archive. Follow the installation steps and then when running nvcc --version in the terminal you should be able to see the installed CUDA version. Then, install PyTorch with CUDA by running:

conda install pytorch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 pytorch-cuda=11.8 -c pytorch -c nvidia

Once you do this, the installation should be complete! In order to verify if everything has been installed successfuly run the following Python script in your tanks Conda environment.

import torch
import stable_baselines3
print(torch.__version__) # 2.1.1
print(torch.version.cuda) # 11.8
print(torch.cuda.is_available()) # True
print(stable_baselines3.common.utils.get_device()) # device(type='cuda')

Running the Program

You can modify the agent rewards in the file gym_tanks/envs/tanks.py. Then, in order to train the agent go to agent_train.py and select the test you want to run changing test_name. In case you want to start training from a certain weights file, specify the number of initial steps of the model you want to load in start_steps. To start the training run:

python agent_train.py

In order to visualize the training results open a new terminal and run:

tensorboard --logdir=logs

In case you don't want to train the agent but just to load a model and see its performance, go to agent_load.py file and select the model you want to load in models_dir. If you want to use the best solution that was reached by us, you can download the weights file from https://drive.google.com/drive/folders/1DgOlP0Pr-otEAJC12qajCUuHw3icMhHb (remember to locate it in models/PPO/Test_Bot folder). Then, run:

python agent_load.py