For our CS 747 - Foundations of Intelligent & Learning Agents course project in Fall 2018, Chinmay Talegaonkar, Dhruv Shah, and I took on the Pommerman Challenge – a test bed for multi-agent learning algorithms.
Pommerman is essentially a clone of the classic Bomberman game: a constantly changing environment with bombs going off, power-ups to collect, and other agents to eliminate. There are two main modes: FFA (four agents battle each other, last one standing wins) and Team (two teams of two agents compete).
We took on the FFA challenge and explored several approaches, including vanilla DQN, rule-based heuristics, and imitation learning. After hitting various walls, we found that the DQfD (Deep Q-learning from Demonstrations) agent performed remarkably well compared to our other approaches – it turns out that showing an agent how humans play before letting it loose is quite effective.
Our experiments and approaches are detailed in the full report: Report
To see our agent in action, watch the video below: