Evolution of learning parameters in a team of mobile agents

Most of everyday life environments are unknown and dynamic. Therefore, the artificial agents living in such environments must adapt their policy based on the environment conditions. In this paper, we consider a team of mobile agents that learns to survive by capturing the active battery packs. In our method, evolution considered metaparameters of an actor-critic Reinforcement Learning algorithm. Results show that after some generations the agents were able to survive and increase the energy level. In addition, the evolved metaparameters helped the agent to adapt much faster during the first stage of life and find a important relation between exploration-exploitation and energy level.