Gaming with AI: A Hybrid Reinforcement Learning, Large Language Model, and Procedural Content Generation Framework for Enhancing Player Engagement and User Experience

Abstract

Most existing game AI research examines individual mechanisms-Dynamic Difficulty Adjustment (DDA), large language model (LLM)-driven NPC dialogue, and Procedural Content Generation (PCG)-in isolation, leaving open the question of how these subsystems interact when evaluated together against a shared user-experience (UX) measure. This paper presents and evaluates a Hybrid AI Gaming Framework that runs three complementary modules asynchronously: (1) a Proximal Policy Optimisation (PPO) reinforcement-learning agent for real-time difficulty calibration, (2) a vector-grounded Memory Repository that anchors LLM-based NPC dialogue to the game's canonical world state, and (3) a deep-learning player-behaviour model that drives PCG-based content variety. The system was deployed in Unreal Engine 5, using a quantised Llama-3-8B-Instruct model served over a Python microservice, with Qdrant as the vector store. A double-blind user study (N = 80) yielded statistically significant improvements across all seven Game Experience Questionnaire (GEQ) dimensions (p < 0.001), with Flow State reaching 3.64/4.0 and Immersion 3.78/4.0. Average session length increased by 42.5%, and Day-14 retention reached 78.4% versus 22.5% for the scripted baseline. The Memory Repository held narrative hallucination to just 0.4% across 25 dialogue turns, and an ablation study confirmed that every architectural component makes a distinct, non-redundant contribution. Together, the results suggest that hybrid, task-differentiated AI architectures yield engagement gains that no single paradigm can match.