A Multi-Agent Framework for Replicating Human Strategic Behaviour via Hypergames

This research examines the intersection of game theory and multi-agent systems, focusing on multi-agent-based simulations of strategic interactions. Combining the practicality of agent-based systems and the conceptual foundations of game theory, we aim to define and ground artificial decision-makers that approximate human behaviour. Human decision-making is context-dependent and shaped by heterogeneous perceptions, which conventional models may fail to capture. Although there are game-theoretic extensions that model uncertainty and heterogeneity to a greater extent, these approaches often lack the pragmatism necessary for dynamic applications that realism-inspired multi-agent simulations entail.

Addressing this gap, our work operationalises hypergames as a practical extension of classical game models. It introduces the Multi-Agent Centralised Hypergames (MACH) framework, a role-based simulation environment governed by a central agent that facilitates, validates, and rationalises gameplay. The framework integrates a novel, hypergame-specific domain-specific language (DSL) that formalises subjective games and extends classical equilibrium concepts into Hypergame Nash Equilibria, enabling the automated reconstruction of belief hierarchies through ``hypergame rationalisation''. Building upon this foundation, we develop a hybrid agent architecture inspired by cognitive theories of human decision-making. We implement both learning-based and large language model-driven reasoning modules, enabling both analytical and natural language representations of strategic reasoning.

Our empirical results illustrate the capabilities of the MACH framework and provide a benchmark of agent performance in prisoner's dilemma and beauty contest games. Our case studies examine the potential autonomy offered by hypergame rationalisation and autoformalisation of game descriptions. Our contributions, which blend the expressiveness of hypergame theory with the practical constraints of multi-agent-based simulations, establish a generalisable foundation for replicating human-like reasoning. We advance both theoretical and practical developments, enhancing the accuracy of representations of human-like strategic behaviour. Our findings enable and inspire promising future research, including an extended DSL that supports logically grounded reasoning and the semantic validation of reasoning.