New model maps how social norms evolve to solve cooperation dilemmas
Stochastic signals and Markov games reveal the fitness landscape of norms.
Social norms offer a decentralized way to resolve classic cooperation dilemmas—like the prisoner's dilemma or public goods games—by prescribing behavior across multiple agents based on stochastic signals. In a new paper on arXiv, Maximilian Puelma Touzel formalizes a “fitness landscape” for such norms using evolutionary game theory. He adapts the rationality criteria framework from Morsky & Akçay (2019) into the more flexible Markov game setting, which is also the foundation of modern reinforcement learning theory. This allows modeling norms where agents interpret imperfect, correlated signals to decide whether to cooperate or defect. The key insight: signals must strike a balance—enough correlation to orchestrate coordination, yet enough uncertainty to discourage exploitation.
The paper derives a general solution for replicator dynamics, showing how norms can spread through a population of rational agents that compare and adopt strategies. By mapping norms over signal and reward spaces, Touzel provides a clear visualization of which norms are evolutionarily stable. This work bridges game theory, multi-agent systems, and social network analysis. For AI developers, it offers a mathematical toolkit for designing agents that can autonomously establish cooperative norms in environments with noisy observations—crucial for applications like autonomous driving, resource allocation, or decentralized AI governance.
- Uses Markov games (common in RL) to generalize prior norm classification, accounting for stochastic signals and rational agent comparisons.
- Identifies a critical balance between signal correlation (for coordination) and uncertainty (to prevent exploitation).
- Provides a replicator dynamics solution showing how norms can emerge and stabilize in a population without central authority.
Why It Matters
Enables design of AI agents that autonomously establish cooperative norms in noisy, decentralized systems.