Collaborative Multi-Agent Perception in Pommerman

The overarching goal of this bachelor's thesis is to find out if approaches from the domain of learnable communication can enable collaborative perception in the Pommerman domain. Existing approaches for the team radio mode with partial observability employ heuristics to exchange information between agents in order to increase their win rate. As the communication bandwidth is very limited (only 6 bits per timestep), manually designing a sophisticated communication protocol is complex. Consequently, existing heuristics are kept simple and usually convey very litte information (e.g. own position and position of one opponent). However, both agents have access to further local information that could be worth sharing (e.g. powerup locations, map tiles, bomb locations and movement..). This thesis investigates if approaches with learnable communication could be used to create communication protocols that allow to convey more information under the same bandwidth constraints.