G. Capi, D. Hironaka, M. Yokota, and K. Doya (Japan)
Reinforcement learning, evolutionary algorithms,surviving robot, metaparameters.
Artificial agents, living in dynamic environments, must adapt their policy based on the environment new conditions. In this paper, we propose a new method, based on combining learning and evolution. In our method, evolution optimizes the metaparameters and initial weight connections of an actor-critic Reinforcement Learning algorithm. The performance is tested in a surviving behavior, where the Cyber Rodent robot has to survive and increase its energy level by capturing the active battery packs. Results show that evolved metaparameters, especially cooling factor, helped the agent to adapt and find the optimal behavior in changing environments. In addition, based on evolved initial weight connections, the agent adapted much faster during the first stage of life. On the other hand, the agent controlled by evolved neural controller applied always the same policy.
Important Links:
Go Back
