Can a payoff be negative?

Yes. Negative values can represent less preferred outcomes or costs under the chosen scale.

Is a zero-sum game one where every payoff is zero?

No. It means the players' payoffs sum to zero for each outcome; one player's gain corresponds to another's loss on that modeled scale.

Is a reward function always the same as a payoff function?

The terms overlap, but reinforcement learning often defines rewards over individual transitions, while a game payoff may summarize a final outcome or accumulated result.

AI Glossary / Graphs and symbolic AI / Intermediate

Payoff function

Definition

A payoff function assigns a numerical value to an outcome or combination of players' actions, representing how desirable it is to a particular player. In game theory, each player can have a different payoff function for the same joint outcome.

What it means

A payoff is not necessarily money. It might represent points, task success or a modeled preference combining several consequences. The function supplies the numerical preferences used by an analysis; it does not automatically describe what a real person actually values.

In a strategic-form game, a player's payoff depends on the action profile: one action chosen by each player. In a game tree, payoffs commonly label terminal outcomes. Expected payoffs combine outcome values with probabilities when uncertainty or randomized strategies are modeled.

This differs from a heuristic evaluation of an unfinished game position. The heuristic estimates future desirability, while the terminal payoff specifies the modeled value of an actual outcome.

How it works

Define the players, their available actions and the outcomes produced by joint choices. For each player i, specify a function u_i that maps an action profile or outcome to a number.

A payoff matrix displays these functions for a finite two-player game. Each cell commonly contains an ordered pair, with the first number for the row player and the second for the column player. Confusing the order changes the incentives inferred from the table.

To evaluate a randomized situation, weight a player's outcome payoffs by the appropriate probabilities and sum them. Expected-utility calculations require a utility scale suitable for preferences over uncertain outcomes; an arbitrary ranking alone does not justify every numerical averaging operation.

The resulting model supports best-response and equilibrium analysis. These conclusions remain conditional on the specified payoffs, available actions, information and behavioral assumptions.

Three practical examples

A game-playing program uses win, draw and loss payoffs to compare terminal outcomes.
A multi-agent simulation gives each participant a payoff function reflecting its modeled objective under joint actions.
A coordination analysis uses a payoff matrix to show how one team's preferred choice depends on another team's choice.

Think of it this way

Think of a scoring rule for a competition. It tells participants how outcomes are valued and can change which tactics look attractive.

The analogy is limited because game-theoretic payoffs may model subjective preferences rather than official scores. Two players need not use the same scale, and a larger numerical payoff for one player is not automatically comparable to a smaller payoff for another. The model must explain what the numbers mean.

Worked example

Consider an invented coordination game. The row player chooses Early or Late, and the column player also chooses Early or Late. Payoff pairs, ordered as (row,column), are: (Early,Early)=(4,3); (Early,Late)=(0,0); (Late,Early)=(1,1); and (Late,Late)=(3,4).

Suppose the row player believes the column player chooses Early with probability 0.6 and Late with probability 0.4. The row player's expected payoff from Early is 0.6×4 + 0.4×0 = 2.4. From Late it is 0.6×1 + 0.4×3 = 1.8. Early is the best response to that stated belief.

If the probability of Early is instead 0.2, the expected payoffs become 0.8 for Early and 2.6 for Late, reversing the choice. This does not change the payoff function; it changes the belief about the other player's behavior.

The numbers are invented utility values, not money or measured preferences.

Illustrative case study

The fictional Cedar Teams group builds a coordination simulation. Its first version treats every participant as valuing completion speed identically, even though one role also bears the cost of rework.

Cedar revisits the payoff definitions with the people represented by the model and compares conclusions under alternative plausible values. It separates uncertainty about preferences from uncertainty about what actions others will choose.

The fictional case illustrates that solving a game accurately cannot repair a poorly specified payoff model. The numerical objective must represent the question the organization actually wants to study.

Common misconceptions and mistakes

Payoff always means financial profit. It can represent other modeled preferences or scores.
The largest cell gives the action to choose regardless of others. A player's outcome can depend on all players' actions.
Every player's payoff uses a common interpersonal scale. That comparison needs additional justification.
Changing beliefs changes the payoff function. Beliefs can alter expected payoff while the underlying outcome values stay fixed.
Any ranking can be averaged meaningfully under uncertainty. Expected-utility analysis requires appropriate utility representation.

Business use

Payoff functions help make incentives explicit in simulations, agent design and strategic analysis. They reveal which outcomes a model rewards and where participants' objectives align or conflict.

For business use, document whether numbers represent money, ordinal preferences, expected utility or an engineered score. Check omitted costs and consequences before interpreting an equilibrium or recommending behavior.

Perform sensitivity analysis when values are uncertain. A conclusion that reverses under small plausible payoff changes should be reported differently from one robust across a broad range. Also distinguish predicting real behavior from prescribing a choice under a mathematical model; the payoff table alone cannot establish how people will act.

Check your understanding

In the worked example, why does the row player change its preferred action when the probability of the column choosing Early changes?

Answer: The payoff values remain fixed, but the probabilities weighting those outcomes change. The expected payoff from each action therefore changes, producing a different best response.

Related terms

Utility function
Expected utility
Zero-sum games
payoff matrix
action profile
best response
Nash equilibrium
zero-sum game
reward function

Learn next

Write a two-player payoff matrix and label whose number appears first. Choose a probability distribution for the other player's action and compute your expected payoff for each choice. Then vary the belief and separately vary a payoff to see why these are different modeling changes.

Probability distribution

Sources

Osborne's strategic-games tutorial defines players, action profiles and payoff functions: https://mjo.osborne.economics.utoronto.ca/index.php/tutorial/index/2/SGA/t

Nau's game-theory materials discuss payoff matrices and mixed-strategy expected utility: https://www.cs.umd.edu/~nau/cmsc421/game-theory.pdf

The Early/Late matrix, expectation calculations and Cedar Teams case are original teaching examples. They do not estimate actual preferences or predict a real organization's behavior.

Previous: Pattern theory Next: PD controller