Modelling Behaviour in Uncertainty: A Simulation Study of Heads-Up Poker.

Modelling human behaviour in situations of incomplete information is a domain with significant implications for the development of AI, empirically understanding facets of psychology, decision theory, sociology, and economics. Due to the suitability of poker to the structure of this particular problem, this study employs the game to propose a model for human behaviour using a probability distribution over the array of possible decisions, incorporating parameters that reflect real characteristics: confidence(⍺), risk appetite(β), and bluff frequency. It uses different combinations of these parameters to run Monte-Carlo simulations of heads-up poker games to understand what makes a particular strategy successful. The simulations revealed that strategies with slightly higher confidence and higher risk appetite yielded the best results in situations closer to reality, while slight underconfidence and lower risk appetite fared better against an "ideal" opponent who has no biases. The proposed model offers a computationally efficient tool for deep learning-based poker AI, with applications in opponent modelling, behavioural economics, and agent-based modelling.