Our Bomberman Agent

This is the setup for a project/competition amongst students to train a machine learning agent for the game Bomberman. Since we particpated in this challenge this repository also contains the code for our agent.

Branches

The branch NN-optim-jonas was used to develop the direct policy learning of the imitation learning since this was independent of the Q-learning part.

The branch NN-optim-convolutionalNet contains in addition to the three-layer fully-connected network the convolutional network and implements the behavioral cloning part of the imitation learning.

The branch DLFD contains an implementation of deep Q learning from demonstrations as suggested here https://arxiv.org/abs/1704.03732. In addition, it allows for generating new expert demonstration data during training similar to the epsilon greedy exploration strategy. This branch was used to create the final agent that we handed in for the tournament.