Task switching in multirobot learning through indirect encoding

摘要: Multirobot domains are a challenge for learning algorithms because they require robots to learn cooperate achieve common goal. The only becomes greater when must perform heterogeneous tasks reach that Multiagent HyperNEAT is neuroevolutionary method (i.e. evolves neural networks) has proven successful in several cooperative multiagent by exploiting the concept of policy geometry, which means policies team members learned as function how relate each other based on canonical starting positions. This paper extends algorithm introducing situational allows agent encode multiple can be switched depending agent's state. demonstrated both simulation and real Khepera III patrol return task, where cover an area home called. Robot teams trained with geometry compared not shown find solutions more consistently also able transfer world.