Must the players repeat the same action?

No. The stage game repeats, while actions can change according to history.

Does a high discount factor mean future rewards matter less?

No. Under the usual convention, a factor closer to one gives future rewards greater weight.

Is a repeated game the same as any sequential decision problem?

No. It specifically models strategic interaction through repeated stage games, rather than merely a single decision-maker acting over time.

AI Glossary / Graphs and symbolic AI / Intermediate

Repeated game

Definition

A repeated game is a game-theoretic model in which the same stage game is played more than once by the participants. Strategies can condition current choices on the history of earlier play. Because today's action may change future responses and rewards, incentives can differ from those in a single isolated encounter.

What it means

The word game refers to strategic interaction, not necessarily entertainment. Each participant chooses actions whose consequences depend partly on what others choose. Repetition adds a future: participants may reward cooperation, respond to departures from an agreement, or learn about each other's behavior.

A repeated-game strategy is therefore more than choosing one action repeatedly. It specifies what to do after possible histories, including histories that do not occur when everyone follows the proposed strategy. The expected duration of interaction and the ability to observe past choices are central parts of the model.

How it works

Specify the stage game's players, available actions, and payoffs. Define how many rounds occur or how continuation is modeled, what each player observes, and how rewards across rounds are combined.

For an indefinitely repeated interaction, future rewards are often discounted by a factor delta between zero and one. A higher factor gives future payoffs more weight. Analyze whether a player can improve their total payoff by departing from a proposed strategy, taking the others' future responses into account.

Credibility matters: a threatened response should itself make sense when the relevant history occurs. Known finite horizons can produce different conclusions from indefinite repetition. Results also change when actions are imperfectly observed or players have incomplete information about one another.

Three practical examples

Multi-agent experiments: compare agents that respond to an opponent's past choices with agents that choose the same action regardless of history.
Shared-resource simulations: study how repeated encounters affect whether participants contribute to maintaining a resource or take immediate benefits without contributing.
Service relationship analysis: use a simplified repeated interaction to examine how future dealings can influence present reliability, while keeping actual contracts and institutional rules separate from the model.

Think of it this way

Think of repeatedly borrowing equipment from the same neighbor. Returning it carefully may affect whether borrowing remains possible later. The analogy captures future consequences, but a formal repeated game needs explicit actions, payoffs, observations, and strategies. Real relationships include emotions, norms, and changing circumstances that a small payoff table may omit.

Worked example

Consider a fictional two-player stage game with actions Cooperate and Defect. Mutual cooperation gives each player 4 points. A lone defector receives 6 while the cooperating player receives 0. Mutual defection gives each player 1. Defection is individually attractive in a single round whichever action the other player chooses.

Assume perfect observation and an infinite discounted horizon. Both players propose: cooperate while every earlier round showed mutual cooperation; after any defection by either player, defect forever. This is a grim-trigger strategy.

If cooperation continues, each player's discounted payoff is 4/(1−delta). Defecting now produces 6 immediately and then 1 each later round, worth 6 + delta/(1−delta). Cooperation is at least as attractive when 4 >= 6−5delta, or delta >= 0.40.

At delta = 0.60, cooperation is worth 10. Defecting now is worth 6 + 0.60/0.40 = 7.5. In the punishment phase, choosing Defect remains a best response to the other player's Defect because 1 exceeds 0. Thus the punishment is credible under these specified strategies and payoffs.

These invented points illustrate incentives, not a prescription for handling mistakes. If an observation error can trigger permanent punishment, the strategy may produce undesirable outcomes despite satisfying the idealized calculation.

Illustrative case study

Fictional case: Juniper Agent Classroom runs repeated resource-sharing games. An initial report counts only total points from one short tournament and declares one strategy universally best.

The instructor asks students to vary the opponent mix, continuation assumptions, and observation noise. Students discover that their original ranking answered a narrower question than they thought. They document how each agent responds after cooperation, defection, and mistaken observations rather than describing agents only as “friendly” or “competitive.”

The lesson is about specifying the interaction and testing robustness. It does not report an executed experiment or claim that one strategy will dominate all realistic settings.

Common misconceptions and mistakes

Repetition does not guarantee cooperation. Players may value the future too little, lack reliable observations, or face incentives that the proposed future responses cannot offset.

Do not transfer infinite-horizon conclusions automatically to a known final round. In the standard finitely repeated prisoner's dilemma with complete information and the usual rationality assumptions, backward induction supports defection throughout.

An equilibrium is also not a guarantee of social desirability or a prediction that people will follow it. It describes mutual incentives within a specified model. Multiple equilibria and differences between model assumptions and actual behavior can remain.

Business use

Repeated-game thinking helps teams identify where a one-period analysis misses future consequences. For AI systems interacting repeatedly, evaluation should include histories, opponent adaptation, and the effects of errors rather than only immediate rewards.

Keep reward definitions explicit. An agent maximizing its own points may undermine a broader service objective. Test changes in the interaction horizon and information available to participants. Use the model to ask clearer questions about incentives; do not treat a stylized equilibrium calculation as evidence about a specific real organization's behavior.

Check your understanding

In the worked example, delta is 0.20. Does the proposed cooperation incentive hold?

Answer: No. Cooperation yields 4/0.80 = 5, while immediate defection followed by punishment yields 6 + 0.20/0.80 = 6.25. The proposed strategy does not deter that departure.

Related terms

Game theory
Strategy
Discount factor
stage game
Nash equilibrium
subgame perfection
prisoner's dilemma
multi-agent learning

Learn next

Write a strategy as a rule for every relevant history, then compare its continuation payoff with a departure. Next, study finite-horizon backward induction and imperfect monitoring. These topics explain why duration and observation assumptions can matter as much as the one-round payoff table.

Sources

MIT OpenCourseWare, Repeated Games chapter: https://ocw.mit.edu/courses/14-12-economic-applications-of-game-theory-fall-2025/mit14_12f12_chapter12_f12.pdf Yale, ECON 159 Lecture 22: https://oyc.yale.edu/economics/econ-159/lecture-22 Primary teaching material was read. Payoffs, calculations, and fictional classroom case are original.

Previous: Repeatability Next: Repeated measures design