Abstract: Traditional C}learning algorithm is based on a single standard of reward, when the environments and the state is changed, the single standard of reward may not be able to adapt to new environments and state in multi agent system(MAS) , instead, it may restrict the learning efficiency. hhis paper proposed a method of multi agent "lcarning algorithm with multi-standard of reward. It adapt well to the changing environment and the state, complete the task in stages, different stages use different standards, so it can quickly complete the stage goal. In this paper, the simulation platform is pursuit problem in threcdimensional world. We increased the difficulty of rounding up and the complexity of the environment and state. Simulation results show that "lcarning algorithm based on multi-standard of reward can flexibly adapt to different environments and state,and efficiently complete learning tasks.

Key words: "learning algorithm, Multi-standard of reward, MAS, Pursuit problem in threcdimensional world