Bayesian nonparametric models characterize instantaneous strategies in a competitive dynamic game

Area of Science:

• Behavioral economics
• Computational neuroscience
• Game theory

Background:

• Traditional game theory models often simplify strategic interactions with fixed turns and limited choices.
• Real-world social behaviors involve complex, coevolving decisions, challenging existing modeling approaches.

Purpose of the Study:

• To quantify dynamic coupling between interacting agents in a strategic game.
• To model human decision-making using reinforcement learning and Gaussian Processes.
• To analyze timing strategies in relation to opponent behavior.

Main Methods:

• Developed a game for human vs. human and human vs. AI competition.
• Applied reinforcement learning with Gaussian Processes to model participant policies and value functions.
• Analyzed decision-making based on game state and opponent identity.

Main Results:

• Quantified instantaneous dynamic coupling between human and artificial agents.
• Identified that high-scoring participants timed directional changes when opponent strategies were weakest.
• Observed less precise move timing in lower-scoring participants.

Conclusions:

• The developed approach allows for multi-timescale analysis of strategic interactions.
• Reinforcement learning with Gaussian Processes effectively models dynamic agent behavior.
• This methodology opens new avenues for experimental paradigms in behavioral analysis.