Autonomously Acquiring a Video Game Agent’s Behavior: Letting Players Feel Like Playing with a Human Player

Abstract

Designing behavior patterns of video game agents (COM players) is a crucial aspect of video game development. While various systems aiming to automatically acquire behavior patterns has been proposed and some have successfully obtained stronger patterns than human players, the obtained behavior patterns looks mechanical. We present herein an autonomous acquisition of video game agent behavior, which emulates the behavior of a human player. Instead of implementing straightforward heuristics, the behavior is acquired using Q-learning, a reinforcement learning, where, biological constraints are imposed. In the experiments using Infinite Mario Bros., we observe that behaviors that imply human behaviors are obtained by imposing sensory error, perceptual and motion delay, and fatigue as biological constraints.