Evolving general cooperation with a Bayesian theory of mind

Area of Science:

• Evolutionary Game Theory
• Artificial Intelligence
• Cognitive Science

Background:

• Cooperation theories often use inflexible models lacking a theory of mind.
• Self-interested agents struggle to achieve collective benefits without understanding others' intentions.

Purpose of the Study:

• To develop and evaluate a novel model of reciprocity incorporating a theory of mind.
• To demonstrate the advantages of inferring others' mental states for fostering cooperation.

Main Methods:

• Developed the Bayesian Reciprocator, a model that values others' payoffs based on inferred cooperation.
• Utilized a probabilistic, generative approach to infer latent beliefs and strategies.
• Tested the model in unique games and classic iterated prisoner's dilemma scenarios.

Main Results:

• The Bayesian Reciprocator sustains cooperation through direct and indirect reciprocity.
• It outperforms existing automata strategies in evolutionary competition.
• Cooperation is sustained across a wider range of environments and noise levels.

Conclusions:

• A theory of mind provides a significant advantage for cooperation in evolutionary game theory.
• The Bayesian Reciprocator offers a more human-like learning mechanism for AI agents.
• This approach facilitates cooperation in diverse and complex interactive environments.