Verifying adaptation of neuro-controlled game opponent by cross validation under supervised and unsupervised player modeling

As a new trend of game development, Artificial Intelligence (AI) used to be implemented by Finite State Machine (FSM) is being replaced by domain-knowledge-free automatic game design approaches. These training methods could achieve excellent performance as FSM and contribute to other improvements: Non Player Character's (NPC) behaviors appear wide differences based on player action, and for multiple opponents in a single game, they can learn to cooperate and form anti-strategy, which could greatly augment entertainment of game. In this paper, by using Pac-man of prey and predator game genre as test-bed and Upper Confidence bound for Trees (ÜCT) algorithm to generate training data for Artificial Neural Network (ANN), we demonstrate adaptation is confirmable with cross validation result under supervised and unsupervised player modeling. As a conclusion, we will clarify how theory developed in this paper could contribute to practical application in online video game and exert excellent game performance.