Architecture

We were able to implement all the descendants of Rainbow DQN except Categorical DQN

• basic DQN
• Double DQN
• Prioritized Experience Replay
• Dueling Network Architectures
• Noisy Nets
• Multi Step Reinforcement Learning
• Categorical DQN

Getting Started with our DQNAgent

Pre-requisites

• Python 3.6.0+ (including pip)
• Docker (only needed for DockerAgent)
• tensorflow 2.6.0
• Keras 2.6.0
• Others are all included in requirements

Installation

• Clone the repository

$ git clone https://github.com/oxFFFF-Q/Project_AI.git

Pip

• Install the pommerman package. This needs to be done every time the code is updated to get the latest modules

$ cd ~/playground
$ pip install -U .

Conda

• Install the pommerman environment.

$ cd ~/playground
$ conda env create -f env.yml
$ conda activate pommerman

• To update the environment

$ conda env update -f env.yml --prune

Launch the agent

We have seperately trained models for player 1 Agent1 and player 3 Agent3. Run main_test.py to test them palying against two SimpleAgent.

Train your agent

A Simple Example

Run main_train.py to train our final DQN model for radio team competition of two SimpleAgent as enemies and a SimpleAgent as teammate.

The training will not automatically stop, but need to be done manully, according to the given out report about the rewards. The paramaters will be recorded every 100 episodes. Run main_save_model.py to save the model. The name of the model is required. The best one is usually among the last few models.

Use other strategies

Select other names for strategy in main_train.py to try other achietectures. Make sure of the consistency of the strategy in main_save_model.py.

Visualize the experiment results

Our experiment results are all stored in data. Run make_image.py to get a visualization of them.

Reference

Our approch get inspired from Sentdex tutorial