Bomberman Reinforcement Learning

Setup for a project/competition amongst students to train a winning Reinforcement Learning agent for the classic game Bomberman.

Setup

First run:

    conda create -n bomberman python=3.8
    conda activate bomberman
    conda install scipy numpy matplotlib
    pip install scikit-learn pygame tqdm

Then follow this tutorial or just execute the following steps:

• Install tensorflow according to the documentation
• Install the CUDA Toolkit v11.2.1
• Install NVIDIA cuDNN v8.1.0 according to the documentation

Also, please create the following directories if you want to start the training process:

    cd agent_code/big_bertha_expert
    mkdir buffers

    cd agent_code/big_bertha_v1
    mkdir buffers

To receive the pickle files for the models uploaded (for pre-initialized replay buffers), please contact the owners of this repository as the files were too large to be uploaded.

Helper Utility

Install following helper tools for development purposes:

    pip install isort
    pip install tensorboard
    conda install pydot
    conda install graphviz

Open a tensorboard after training with:

    cd agent_code/big_bertha_v1
    tensorboard --logdir tensorboard_logs

Training process

The training process of the model is based on Deep Q-learning from demonstrations. Therefore two agents exists:

1. big_bertha_expert (similar to the rule based agent) -> Our network is pretrained with the filled experience replay buffer from the expert
2. big_bertha_v1 (our DQN) -> Trained network from 1. is trained with normal Reinforcement Learning

Step 1. is part of the Imitation Learning and step 2. is standard Reinforcement Learning.