Hybrid Intelligence in Evolving Games: Automated Rule Design, Strategy Evolution, and Evaluation Optimisation for Intelligent Societies

Abstract

Adaptive rule evolution is a fundamental challenge in artificial life and multi-agent systems research, as traditional game environments rely on static, human-designed rules that cannot adapt to emergent behaviours. This dissertation addresses this limitation by exploring automated, dynamic rule creation mechanisms in multi-agent digital environments. It first introduces a structured Strategy-Evaluation-Rule (SER) framework for rule generation that formalises the interplay between agent strategies, evaluative feedback, and rule adaptation. Unlike prior approaches to game rule design, SER does not rely on any prepared datasets or domain knowledge; instead, it generates rules on the fly and refines them through iterative self-play evaluation. The SER framework is implemented in two games, Maze Run and Trust Evolution, demonstrating its effectiveness in driving emergent, complex agent behaviours across disparate domains.

Building on this foundation, the thesis presents the Triadic Reciprocal Dynamics (TRD) system, which establishes a novel closed-loop paradigm linking rule creation, strategy evolution, and performance evaluation. TRD is instantiated as a multi-agent game environment where AI, NPCs, and human players participate together. It employs a neural rule designer and an automated evaluator to continuously generate new game rules and assess their impact on evolving strategies in real time.

Furthermore, principles from Flow Theory are incorporated into the SER framework to enable flow-driven rule design, ensuring that gameplay remains engaging. This extension features dynamic difficulty adjustment (DDA), which dynamically tunes challenge levels in real-time, and a dual-reward scheme that balances extrinsic rewards (e.g., points or achievements) with intrinsic signals of player engagement. A real-time flow visualisation interface is also introduced to monitor players' flow states and guide on-the-fly adjustments to maintain an optimal flow experience. Together, these contributions advance the state of the art in automated game design and adaptive multi-agent systems by enabling rules that evolve autonomously, exemplifying a novel paradigm of dynamic rule evolution and laying the groundwork for future research in lifelike, self-evolving multi-agent environments.