Reinforcement-learning agents with different temperature parameters explain the variety of human action- selection behavior in a Markov decision process task

摘要

We investigated the characteristics of the human action-selection in performing a Markov decision process (MDP) task, and compared them to those of reinforcement-learning (RL) agents. The behavior of human participants was roughly classified into two qualitatively different types. On the other hand, surprisingly, the variety of human behavior could be explained simply by a single parameter of the degree of randomness (i.e., the temperature parameter) in the action-selection rules of the RL agents. This result implies that the various behaviors of human action-selection may be determined by a simple mechanism in the brain.