Pommerman- Reinforcement Learning Project
As the course project for CS 747 - Foundations of Intelligent & Learning Agents, Fall 2018, I teamed up with Chinmay Talegaonkar and Dhruv Shah to take on the Pommerman Challenge.
The Pommerman Challenge essentially serves as a test bed for multi-agent learning algorithms. It is essentially a clone of the popular Bomberman game, providing a constantly changing environment with bombs going off, power-ups to be collected and other agents to be eliminated. There are mainly two modes to the game, FFA- where each of the four agents battle against each other and the last one to survive wins, Team- where there are two teams of two agents which compete against each other.
We took on the FFA challenge, trying a few different approaches. After facing a few hurdles, we found that the DQfD agent performed remarkably well as against our other approaches.
Our experiments and approaches are detailed here- Report
To see our agent in action, watch the video below-