Two-stage training algorithm for AI robot soccer

In multi-agent reinforcement learning, the cooperative learning behavior of agents is very important. field heterogeneous among different types in a group pursued. Learning joint-action set during centralized training an attractive way to obtain such behavior; however, this method brings limited performance with agents. To improve training, two-stage which allows multiple roles proposed. During two processes are conducted series. One stages attempt each agent according its role, aiming at maximization individual role rewards. The other for as whole make them learn behaviors while attempting maximize shared collective rewards, e.g. , team Because these series every time step, can how rewards and simultaneously. proposed applied 5 versus AI robot soccer validation. experiments performed environment using Webots simulation software. Simulation results show that train robots effectively, achieving higher compared three approaches be used solve problems multi-agent. Quantitatively, trained by improves score concede rate 5% 30% when teams matches against evaluation teams.