Discovering Strategic Behaviour of Multi-Agent Systems in Adversary Settings

Abstract

Can specific behaviour strategies be induced from low-level observations of two adversary groups of agents with limited domain knowledge? This paper presents a domain-independent Multi-Agent Strategy Discovering Algorithm (MASDA), which discovers strategic behaviour patterns of a group of agents under the described conditions. The algorithm represents the observed multi-agent activity as a graph, where graph connections correspond to performed actions and graph nodes correspond to environment states at action starts. Based on such data representation, the algorithm applies hierarchical clustering and rule induction to extract and describe strategic behaviour. The discovered strategic behaviour is represented visually as graph paths and symbolically as rules. MASDA was evaluated on RoboCup. Both soccer experts and quantitative evaluation confirmed the relevance of the discovered behaviour patterns.