AI Glossary / Graphs and symbolic AI / Beginner

Game theory

Definition

Game theory studies situations in which the outcome for each decision-maker depends on choices made by others as well as their own. It models strategic interaction through players, available strategies, information and preferences over possible outcomes.

What it means

A game in this sense need not be recreational. It can describe competing automated agents, shared-resource decisions, bargaining or coordination. The central feature is interdependence: the best choice may change when another participant changes their behavior.

Players can be people, organizations or software agents. Payoffs represent what the model says they value; they need not be money. Strategies describe choices or, in a sequential setting, plans for different circumstances.

Game theory does not assume every situation is a contest with one winner. Some interactions permit mutual gains, some involve conflicting interests, and many contain both. Its conclusions depend on the modeled choices, information and objectives being appropriate.

How it works

Begin by identifying who chooses, what each can do and when decisions occur. Specify what each participant observes and how each evaluates the resulting outcome. A payoff table can represent a small simultaneous-choice game; a tree can represent choices over time.

Analysis may compare best responses, identify dominant strategies or find an equilibrium. In a Nash equilibrium, no player can improve their own payoff by changing strategy alone while the others keep theirs fixed. That condition does not imply fairness or the best combined outcome.

Other concepts address cooperation, repeated interaction, incomplete information or the design of incentives. A model can reveal tensions and dependencies, but it is not a prediction machine that automatically knows real participants' preferences or reasoning.

Three practical examples

• Designers of software agents analyze how independently chosen resource policies interact when several agents share a limited service.
• A coordination exercise examines whether teams have incentives to choose compatible standards or schedules.
• A simulation explores competing strategies in a board game or another explicitly modeled multiagent environment.

Think of it this way

Think of choosing a meeting location when each person cares both about convenience and about being together. Your preferred choice depends on where the others intend to go.

The analogy shows strategic dependence. It does not mean the participants know everyone's preferences exactly, calculate perfectly or will settle on a single outcome without communication.

Worked example

Finding an equilibrium that is not jointly best

Consider two fictional automated jobs sharing a simulated resource. Each chooses a restrained policy R or an aggressive policy A. Define payoffs as follows: both restrained receive (3, 3); if the first is aggressive and the second restrained, they receive (5, 0); the reverse gives (0, 5); both aggressive receive (1, 1).

For the first job, if the second chooses R, choosing A gives 5 rather than 3. If the second chooses A, choosing A gives 1 rather than 0. Thus A is strictly better for the first job against either choice. The same reasoning applies to the second.

The pair (A, A) is therefore a Nash equilibrium: neither can improve by switching alone. Yet (R, R) gives both a higher payoff, 3 rather than 1. Individual incentives in this specified game do not deliver the jointly better outcome.

These invented payoffs illustrate an incentive structure. They are not measured processing speeds, and adding enforcement, repetition or communication would require a changed model rather than an automatic claim that cooperation will emerge.

Illustrative case study

The fictional Cedar Agents team observes two scheduling agents repeatedly requesting the same shared capacity. Each agent was evaluated separately and rewarded for completing its own work quickly. Joint tests reveal undesirable interaction.

The team constructs a small game model to expose how local objectives could encourage competing requests. It then reviews the actual resource mechanism and explores changes to allocation rules and objectives. Each proposed change is evaluated in the revised setting instead of assuming that a label such as cooperation fixes the incentives.

Cedar also compares the model with recorded behavior because the agents have timing constraints and limited information. The case illustrates how strategic analysis can frame a design problem without claiming the payoff table is a complete account of production behavior.

Common misconceptions and mistakes

• Every game is zero-sum. Many games allow both participants to benefit or suffer together.
• Equilibrium means the best outcome for everyone. It describes a stability condition relative to unilateral changes, not collective optimality.
• Rationality means selfishness in ordinary language. A modeled payoff can include concern for others, rules or long-term objectives.
• An equilibrium calculation guarantees real behavior. Missing information, learning, mistakes and model misspecification can change what actually happens.

Business use

For business and AI projects, game theory is useful when interacting decisions and incentives matter. It can help reveal why optimizing each component separately produces poor overall behavior, or why a proposed policy invites a response that undermines its purpose.

Use models to make assumptions discussable: whose objective is represented, what actions are available and what information is observed. Test conclusions against alternative payoff estimates and realistic constraints. The value often lies in identifying dependencies and incentive problems rather than producing a single supposedly certain forecast.

People also ask

Is a strategy the same as one action?

In a simple one-shot game it may be. In a sequential game, a strategy specifies what to do in different decision situations.

Does every game have a pure-strategy Nash equilibrium?

No. Finite games have an equilibrium when mixed strategies are allowed, but some have no equilibrium using only pure strategies.

What is a mixed strategy?

It is a probability distribution over a player's pure strategies, representing randomized choice under the model.

Check your understanding

Answer: At (A, A) in the worked example, one job switches to R while the other stays at A.

Does the switching job improve its payoff?

Answer: No. Its payoff falls from 1 to 0. This supports the equilibrium condition even though both would prefer the outcome where both choose R.

Related terms

• Dominant strategy
• Mixed strategy
• Zero-sum games
• Game tree
• Mechanism design
• Multiagent systems
• Nash equilibrium
• best response
• zero-sum game
• multiagent system

Learn next

Change the worked game's payoff for being aggressive against a restrained partner from 5 to 2. Recheck best responses against both partner choices rather than assuming the earlier equilibrium reasoning still applies unchanged.

More in Graphs and symbolic AI

• G-set - The G-set is the general boundary of a version space in concept learning.
• Gain parameter - A gain parameter is a coefficient that scales a signal or a component of a system's response.
• Gain ratio - Gain ratio is a decision-tree split criterion that divides information gain by the split information of the candidate test.
• Gaussian process - A Gaussian process is a collection of random variables for which every finite collection has a joint Gaussian distribution.
• Generalization hierarchy - A generalization hierarchy organizes concepts or descriptions from more specific to more general.
• Generalization loss - Generalization loss is a model's expected loss on examples drawn from the target data distribution.

Sources

Nash's original paper establishes equilibrium existence for finite games with mixed strategies; publisher-indexed primary text was reviewed, while direct page access failed: https://doi.org/10.1073/pnas.36.1.48

Rubinstein's author-provided notes define strategic games and equilibrium: https://arielrubinstein.org/econt2/lectures/G01.pdf

AIMA's author materials connect strategies and game representations to AI search: https://aima.cs.berkeley.edu/4th-ed/slides-pdf/chapter06.pdf

The resource game and Cedar Agents case are original illustrations, not empirical or prescriptive claims about a real organization.

Previous: Galerkin approximation Next: Game tree