Interactive project · Roth–Erev reinforcement learning

Where does fairness come from?

Two players. No rules about right or wrong. Each one simply repeats whatever paid off last time. Run the simulation and watch norms like trust, fairness and “that’s mine” show up on their own, or fail to.

Round 1

Round 40

Round 600

One run of Divide the Dollar. Each square is a pair of demands; darker means more likely. Choices start out spread evenly and end up concentrated on the fair ½–½ split.

How the players learn

The model comes from Roth and Erev’s work on how people actually learn in repeated games. It’s deliberately simple: no foresight and no morality, just habits that get stronger when they pay.

  1. Play

    Every round, each player picks a strategy. Each strategy has a weight, and heavier strategies get picked more often.

  2. Reinforce

    Whatever a player earned that round is added to the weight of the strategy they used. Success is sticky.

  3. Forget and slip

    Old weights fade a little each round (forgetting, δ). Now and then a player picks at random by mistake (ε).

weight ← (1 − δ) × weight + payoff

Run a world

Pick a game, set the dials, press play. The grid shows how likely each pair of choices is at the current round.

Divide the Dollar

Split a pie between two players. Each asks for a share. If the two requests fit, both get what they asked for; if together they ask for too much, nobody gets anything.

Payoffs row player · column player

Row strategyAsk ⅓Ask ½Ask ⅔
Ask ⅓2, 22, 32, 4
Ask ½3, 23, 30, 0
Ask ⅔4, 20, 00, 0
1000
0.010

How fast old habits fade. Higher means the players react more to recent rounds.

2%

Chance a round is played at random. Mistakes can knock players out of a rut.

Round 1000 / 1000

Settled

Fair: ½ and ½

Row player

Ask ⅓ 3%Ask ½ 97%Ask ⅔ 0%

Column player

Ask ⅓ 1%Ask ½ 99%Ask ⅔ 0%

Now run 200 worlds of Divide the Dollar

One run is a story. Many runs show the odds. Same game, same dials, 200 different strokes of early luck. Here’s where they ended up.

  • Fair: ½ and ½47%
  • Row takes ⅔, column ⅓25%
  • Column takes ⅔, row ⅓23%
  • Still undecided at the end6%
Good for bothLopsided conventionSafe but meagerBad for both

What it suggests about human ethics

Philosophers like Brian Skyrms use models like this to ask how a social contract could grow out of ordinary, self-interested learning instead of being designed by anyone.

Norms can grow from habit

Nobody in the model is told to trust or to share. Stable conventions still form, because people keep doing what has worked for them.

History decides a lot

The same rules can lead to different outcomes. Luck in the first few rounds decides which norm takes hold, and an unfair norm can last just as long as a fair one.

Fairness pulls hardest

When two players split something, the 50/50 split is where they land most often. Equal treatment emerges without being imposed.

Trust is the hard part

Cooperating would make everyone better off, yet it’s fragile. Caution or self-interest usually wins unless something else helps, like reputation, repeat partners or institutions.

What this doesn’t show

The model explains how a norm could arise. It doesn’t tell you which norm is right. Real people also have language, culture, empathy and punishment, and none of that is in this model.

Further reading

Roth & Erev, “Learning in Extensive-Form Games” (1995) · Skyrms, Evolution of the Social Contract (1996) · Skyrms, The Stag Hunt and the Evolution of Social Structure (2004)