Dual-Objective Adaptive Adversary Design with Reinforcement Learning for Dynamic Difficulty Adjustment in a Turn-Based Tactical Game

This creative component investigates dual-objective adaptive adversary design for dynamic difficulty adjustment in a turn-based tactical game inspired by Into the Breach. The project aims to build an adversarial system that adapts to player behavior in real time while balancing two competing policies: increasing challenge and maintaining engagement. Rather than designing an opponent that only maximizes player loss, the work frames difficulty adjustment as a control problem in which the adversary must apply pressure without creating frustration or leading to a repetitive, unfair gameplay. The proposed framework uses a hybrid architecture with two decision layers. A director-level controller selects macro-level parameters such as spawn composition, pacing mode, target priority, and aggression, while a tactical controller determines per-enemy local behavior. The project explores reinforcement learning and adaptive decision-making methods including PPO, A2C, DQN, QRDQN, TRPO, and contextual bandits, along with memory architectures such as MLP, LSTM, and GTrXL. Evaluation is based on both challenge-oriented and engagement-oriented reward signals, including pressure, threat, damage, close calls, frustration, and flow-style metrics. The final system is deployed as a FastAPI server for real-time communication with the game engine, providing a practical and extensible foundation for adaptive adversarial gameplay in rogue-like games. The resulting system provides a foundation for adaptive adversarial gameplay modeling in tactical games. 

Committee: Simanta Mitra (major professor)