Using suitable action selection rule in reinforcement learning

Reinforcement learning under perceptual aliasing causes the following serious trade-off on action selection roles. When an agent decides its action by a deterministic selection rule according to action values, acquired policies may include infinite loops. On the other hand, while a stochastic selection rule can avoid such infinite loops, it reduces the efficiency of the acquired policy. To resolve this trade-off, we propose a method called "adaptive meta-selection", by which agents can determine a better selection rule between deterministic and stochastic ones for each state. Results of experiments in Pursuit Game show that proposed method enables an agent to acquire a policy that has efficient performance while excluding infinite loops.