Towards Setplays Learning in a Multiagent Robotic Soccer Team

Abstract

This work proposes a novel machine learning strategy to build new predefined cooperative plans in multiagent systems. State machines describe these plans, named setplays. Each state represents the current behavior of each agent following a setplay. In turn, a setplay requires conditions to define the transitions to other states and at least one final state. A final state may represent success or failure as a result of a setplay execution. We present a proposal to learn, from a human specialist knowledge, a dataset to enable agents to learn new setplays. Preliminary results from dataset exploration experiments are presented and future work is described.