🧬 biology

The fitness landscape of social norms in social dilemmas

This paper extends the evolutionary game theory of social norms from matrix games to the more general Markov game setting, clarifying the underlying mechanics of how norms evolve through rational signal interpretation and providing a general solution for the replicator dynamics that govern their emergence in social dilemmas.

Original authors: Maximilian Puelma Touzel

Published 2026-05-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Maximilian Puelma Touzel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a world full of people trying to get along, but they are constantly stuck in situations where doing what's best for themselves hurts the group, and doing what's best for the group feels risky for the individual. The paper calls these "social dilemmas." A classic example is the Game of Chicken: two cars speeding toward each other. If both swerve, they look silly but survive. If one swerves and the other doesn't, the swerver looks like a coward, but the other wins. If neither swerves, they crash.

The paper asks: How do people figure out how to coordinate without a boss telling them what to do?

Here is the breakdown of the paper's ideas using simple analogies:

1. The Problem: Guessing Games

Usually, in game theory, we assume everyone is a "rational calculator" trying to beat the other person. But if everyone tries to outsmart everyone else, you often end up in a crash (or a bad outcome).

The authors suggest that instead of just guessing what the other person will do, people use signals from the environment. Think of a traffic light. The light doesn't force you to stop; it just suggests it. If you trust the light, you stop. If you trust that the other driver also trusts the light, you both stop safely. This is called a Correlated Equilibrium. It's a way of coordinating where everyone follows a shared rule based on a signal, rather than trying to outsmart each other.

2. The Solution: Social Norms as "Rulebooks"

The paper introduces the idea of a Social Norm. Think of a norm not just as a rule, but as a two-part belief system:

The Prescription: "When I see a red light, I stop." (What I do).
The Description: "When you see a red light, you stop." (What I believe you will do).

For a norm to work, it has to be rational. If the norm says "Stop at red," but you realize that "Going at red" gives you a better reward (maybe you're in a hurry and the other car is far away), you will break the norm. The paper maps out exactly when these "rulebooks" make sense and when they fall apart.

3. The Evolution: Survival of the Fittest Rules

How do good norms appear? The paper uses a concept called Replicator Dynamics. Imagine a giant room full of people.

Some people follow the "Stop at Red" rule.
Some follow the "Go at Red" rule.
Some follow the "Always Go" rule.

Every day, people pair up and play the game. If you follow a rule that gets you a high reward (you survive the intersection), you are "fit." If you follow a rule that gets you a crash, you are "unfit."

The paper shows that over time, the people with the "unfit" rules drop out, and the people with the "fit" rules (the ones that coordinate well) take over the room. It's like natural selection, but instead of genes, ideas and rules are evolving.

4. The "Chicken" Experiment

The authors tested this specifically with the "Game of Chicken." They created a mathematical map (a phase diagram) to see under what conditions a "Signal-Following Norm" (like obeying a traffic light) would win.

The Result: They found that if the temptation to "go" (be greedy) is too high, the signal breaks down, and people crash. But if the temptation is manageable, the "Signal-Following" norm becomes the dominant strategy.
The Surprise: The "Signal-Following" norm (where you do exactly what the signal says) is often much better for everyone than the "Nash Equilibrium" (the standard game theory solution where everyone plays it safe but still risks a crash). In their simulations, following the signal could make the group up to 1.8 times better off than the standard "rational" approach.

5. The Big Picture

The paper argues that we don't need a central authority to solve social problems. If we can create stochastic signals (random but correlated hints in the environment) and if the "rules" for following those signals are rational, society can naturally evolve toward a state where everyone coordinates perfectly.

In short: The paper is a mathematical proof that if we agree to follow simple, shared signals (like traffic lights or social cues), and if those signals are designed right, we can evolve a society where we avoid crashes and get along, even when we are all acting in our own self-interest. It's about how "rules of the road" emerge naturally from the bottom up.

Technical Summary: The Fitness Landscape of Social Norms in Social Dilemmas

Problem Statement
Social norms function as a coordination mechanism to resolve social dilemmas among multiple agents. While evolutionary game theory has traditionally modeled how rational agents compare and adopt norms to reach dominant population states, existing frameworks often rely on matrix games and specific rationality criteria (e.g., Morsky and Akçay [2019]). The paper addresses the need to clarify the fundamental ingredients of this theory and extend it to the more general Markov game setting common in multi-agent reinforcement learning (MARL). Specifically, the work investigates how decentralized, wide adoption of norms can be achieved through the interpretation of stochastic environmental signals, balancing the need for correlation to orchestrate mutual benefit against the disincentivizing uncertainty required to prevent exploitation.

Methodology
The authors formulate a minimal version of social norm theory using the notation of reinforcement learning within a two-player, partially observable Markov game setting.

Environment Model: The environment is defined by a state space $S$ and observation spaces $O, O'$ . State transitions are action-independent, and rewards depend solely on actions, not states. The environment is represented by the joint observation distribution $P_{oo'}$ .
Norm Definition: A social norm is formalized as a belief pair $(P_{a|o}, D_{a'|o'})$ , where $P_{a|o}$ is the prescribed policy (how the agent acts given an observation) and $D_{a'|o'}$ is the described policy (how the agent believes the opponent acts). Norms are classified by validity: null (suboptimal), rational (internally valid), empirically validatable (externally valid), consistent (self-validating), and evolutionarily stable (ES).
Correlated Equilibria: The paper posits that joint strategies adopting norms consistent with optimal single-player strategies satisfy a correlated equilibrium condition rather than a Nash equilibrium. A correlated equilibrium is defined as an observation ensemble and a pair of deterministic policies that are best responses to the expected behavior of the other given their observations.
Case Study: The theory is applied to the "Game of Chicken" (an anti-coordination dilemma with parameters $T > B > S > P$ ). The authors parametrize the reward matrix $R$ and the joint observation distribution $P_{oo'}$ using a coordination potential $b$ and a coordination challenge level $L$ .
Dynamics: The emergence of norms is modeled using replicator dynamics, a mean-field description of selection where agents in pairs compare payoffs and adopt the norm with the higher payoff. The dynamics are analyzed via the Jacobian of the flow map to determine stability and attractor basins.

Key Contributions

Generalized Formulation: The paper translates social norm theory from matrix games to the Markov game setting, explicitly mapping norms over signal and reward spaces.
Complete Characterization of Chicken: The authors provide a full characterization of the Game of Chicken family, deriving the rationality conditions for a "signal-following norm" (e.g., stop on red, go on green) across the parameter space of coordination potential ( $b$ ) and temptation ( $L$ ).
Fitness Landscape Analysis: The work constructs the fitness matrix $\Gamma$ for all possible norms in the Game of Chicken (16 total, reduced to 12 rational ones in the specific case studied) and analyzes the resulting replicator dynamics.
Stability Analysis: The paper derives a general solution for the Jacobian of replicator dynamics at single-norm fixed points and applies it to show how the stability of the signal-following norm depends on the game parameters.

Results

Rationality Conditions: The signal-following norm is rational only within a specific region of the $(b, L)$ parameter space. As the coordination challenge ( $L$ ) increases, the maximum coordination potential ( $b$ ) that allows for rational cooperation decreases.
Performance: In the region where the signal-following norm is rational, the average reward significantly exceeds that of the best mixed-strategy Nash equilibrium (up to 1.8 times better in the analyzed case).
Dynamics and Stability:
- The "always go" norm is always unstable.
- The mixed Nash strategy is a stable fixed point but possesses a negligibly small attractor basin; small perturbations ( $10^{-3}$ ) are sufficient to move the system away from this equilibrium.
- The signal-following norm acts as a dominant attractor in a large portion of the state space when coordination potential is high.
- As coordination potential decreases, the system transitions to a multistable regime where the "anti-signal" norm (doing the opposite of the signal) becomes stable, sharing the attractor basin with the mixed Nash strategy.

Significance
The paper claims to provide a detailed exposition of the underlying mechanics of how correlated equilibria emerge endogenously through the evolution of social norms. By bridging evolutionary game theory, sociology, and multi-agent reinforcement learning, the work motivates the intermingling of empirical and theoretical traditions. It demonstrates that social norms can be understood as a mechanism for agents to achieve correlated equilibria, offering a rational resolution to social dilemmas that outperforms standard Nash equilibria. The analysis of the fitness landscape and replicator dynamics offers a theoretical foundation for understanding how specific norms (like traffic rules) can stabilize and dominate in populations of learning agents.