A Study on Reinforcement-learning Agents with Personality through the Implementation of Character Parameters
Daichi Ando, Shino Iwashita · 2020
We aim to achieve characteristic action in reinforcement-learning agents. Agents with different personality parameters move through a maze with the aim of reaching a goal while obtaining rewards. Three types of parameters are defined: A: Prioritize additional rewards, S: Prioritize goal, and G: Prioritize search rewards. We trained a reinforcement-learning agent with parameters that correspond to a character. Statistical information of the learning process of agents were obtained by using TensorBoard and compared. As a result, the agents did not show any personality traits in their behavior. However, we confirmed that the learning process changed depending on the parameter setting pattern.