Creating Adaptive Adversaries for Fighting Games Using Deep Reinforcement Learning

Deep Neural Networks and Reinforcement Learning have achieved remarkable success in artificial intelligence, surpassing human-level performance in various complex environments. However, creating truly adaptive agents for dynamic environments remains a challenge. Most existing AI agents are fine-tuned to specific strategies, limiting their ability to adapt mid-simulation. This paper presents a novel approach to developing adaptive adversaries in fighting games using Deep Reinforcement Learning. We utilize Street Fighter II as our testbed, training agents with Proximal Policy Optimization and actor-critic reinforcement learning. Our methodology incorporates reward shaping to create distinct offensive and defensive strategies, with agents trained against scripted opponents mimicking these approaches. A heuristic-based decision system enables real-time strategy switching based on game state analysis. Experimental results demonstrate that our adaptive agent achieves a 58% win rate against in-game AI and significantly outperforms single-strategy models, with win rates of 59%, 88%, and 94% against defensive, aggressive, and generic models, respectively. This research presents a framework for developing AI agents capable of dynamically adapting their strategies, with applicability across a range of competitive gaming environments.