Presentation Information
[4Yin-A-03]Multifaceted Evaluation of Adaptive Fighting Game AI Scaling to Diverse Player Skill Levels
〇KEN YAMANE1, Chao Tang1, Taketo Yoshida1, Keigo Tokuda1, Takumi Kakee1 (1. Teikyo University)
Keywords:
Adaptive Game AI,Dynamic Difficulty Adjustment,Dramatic Tension Management,Selective Desensitization Neural Netowork,Pattern Coding
AI characters that compete against human players are crucially important elements in fighting games. To entertain players, these AIs must adjust their behavior rapidly to accommodate diverse strategies and skill levels. We earlier proposed a reinforcement learning AI that uses a selective desensitization neural network as an action-value function approximator. However, its effectiveness had not been verified sufficiently. For this study, we conducted multifaceted evaluation of the proposed AI. Experimentally obtained results against 15 different AI opponents demonstrated that the system not only adjusted win rates of 45% - 53% but also orchestrated close matches in 34% - 85% of the rounds. Furthermore, subjective evaluations confirmed that the AI's intentional errors were perceived as natural and difficult for participants to detect. These findings suggest that the proposed AI has potential to exceed a mere balance of win rates, synchronizing itself with diverse player skill levels to create natural and engaging close-match experiences.