Evolving Action Abstractions for Real-Time Planning in Extensive-Form Games

摘要: A key challenge for planning systems in real-time multiagent domains is to search large action spaces decide an agent’s next action. Previous works showed that handcrafted abstractions allow focus their on a subset of promising actions. In this paper we show the problem generating can be cast as selecting pure strategies from pool options. We model selection two-player game which strategy set players powerset options— call game. then present evolutionary algorithm solving such Empirical results small matches µRTS our approach able converge Nash equilibrium Also, larger algorithms using derived by are substantially outperform all state-of-the-art tested.