An Autonomous Emotional Virtual Character: An Approach with Deep and Goal-Parameterized Reinforcement Learning

Abstract

We have developed an autonomous virtual character guided by emotions. The agent is a virtual character who lives in a three-dimensional maze world. Results show that emotion drivers can induce the behavior of a trained agent. Our approach is a case of goal parameterized reinforcement learning, which creates the proper conditioning between emotion drivers and a set of goals that determine the behavioral profile of a virtual character. We train agents who can randomly assume these goals while maximizing a reward function based on intrinsic and extrinsic motivations. A mapping between motivation and emotion was carried out. So, the agent learned a behavior profile as a training goal. The developed approach was integrated with the Advantage Actor-Critic (A3C) algorithm. Experiments showed that this approach produces behaviors consistent with the goals given to agents, and has potential for the development of believable virtual characters.