Mastering Stochastic Partially Observable Asymmetrical Multi-Agent Game with Rolling Horizon Evolution Algorithm

Hunting of the plark is a complex adversarial multi-agent, asymmetrical, stochastic, and partially observable war game. An aircraft player has the objective to find and destroy a hidden submarine, and the submarine player has the objective of escaping the faster aircraft. This paper demonstrates a statistical forward planning rolling horizon evolutionary algorithm agent, that models enemy behavior using heuristics, and masters the game outperforming other agents including deep reinforcement learning, heuristic, and human players.