Does every player need private information?

No. Some players can have a single type or no private payoff information while others have several possible types.

Must types be independent?

No. Correlation is allowed and can materially affect conditional beliefs.

Is a Bayesian game an equilibrium?

No. It describes the strategic environment; equilibrium is a solution concept applied to that environment.

AI Glossary / Statistics and probability / Intermediate

Bayesian game

Definition

A Bayesian game is a game-theoretic model in which players have incomplete information about relevant features of the interaction, such as another player's costs, valuations, or information. Private information is represented by types, and beliefs describe uncertainty about those types. Strategies specify actions as functions of the information available to each player.

What it means

Bayesian games add uncertainty about others' information or preferences to strategic reasoning.

A supplier might privately know its cost, while a buyer knows only possible costs and their probabilities. The buyer considers how different supplier types behave.

The model is distinct from a Bayesian network, which represents probabilistic relationships, and from a Bayes classifier, which chooses labels. A Bayesian game explicitly includes strategic players, actions, and payoffs. The word Bayesian describes the treatment of private information and beliefs; it does not imply that the participants cooperate.

How it works

Specify the players, their action sets, their type sets, and payoff functions. Payoffs can depend on actions and types. In a standard common-prior formulation, also specify a joint probability distribution over the type profiles.

The model can be pictured as nature drawing the types before play. Each player observes its own type and possibly other signals, but generally does not observe the entire type profile. This is a representation of information, not necessarily a literal random draw performed in the real setting.

A pure strategy maps each possible own type to an action; a mixed strategy allows randomization. Expected payoffs average over the other players' types using beliefs conditional on the player's information. If types are correlated, learning one's own type changes beliefs about the others.

A solution concept, such as Bayes-Nash equilibrium, is then applied to the game. Defining a Bayesian game does not itself solve it or establish a unique prediction. Dynamic interactions may require a richer description of information timing and beliefs at later stages, together with an appropriate sequential solution concept.

Three practical examples

Procurement: represent suppliers' private production costs and a buyer's uncertainty when assessing the incentives created by bidding rules.
Digital marketplaces: study how participants with different private valuations respond to pricing or allocation mechanisms.
Multi-agent AI: model agents whose private objectives or resource limits affect strategic choices, while testing whether the assumed information structure matches deployment.

Think of it this way

Think of players privately receiving cards that change the rewards of available moves. Everyone knows the rules and possible cards, but not others' cards. Bayesian games describe these probabilities and type-dependent plans.

Worked example

Consider two fictional firms deciding whether to enter a market. Entry produces gross revenue 6 when alone and 3 when both enter, minus the entrant's own cost. Staying out gives zero. Each firm's private cost is either low, equal to 2, or high, equal to 5.

Specify the joint type probabilities as follows: both low, 0.40; both high, 0.40; first low and second high, 0.10; first high and second low, 0.10. Each firm is low-cost with marginal probability 0.50, but the types are correlated.

A low-cost firm assigns probability 0.40 / 0.50 = 0.80 to its rival being low-cost. A high-cost firm assigns probability 0.10 / 0.50 = 0.20 to its rival being low-cost.

Suppose the rival's proposed strategy enters only when low-cost. A low-cost firm's expected entry payoff is 0.80 × 3 + 0.20 × 6 - 2 = 1.6. A high-cost firm's expected entry payoff is 0.20 × 3 + 0.80 × 6 - 5 = 0.4.

Both prefer entry to zero against that proposed rival strategy. Therefore, the symmetric proposal that only low-cost types enter is not an equilibrium in this game. The example specifies a valid Bayesian game and diagnoses one failed candidate strategy; it does not claim to solve all its equilibria.

Illustrative case study

Fictional case: Elm Market simulates suppliers entering a short-term sourcing event. Its first model assumes suppliers' costs are independent. Operations staff point out that shared input conditions make high costs more likely to occur together.

The analysts revise the joint type distribution and recalculate each supplier type's beliefs. Even when the marginal cost probabilities stay the same, the information contained in a supplier's own cost changes its expectations about competitors.

The team records this distinction and tests sensitivity to the cost relationship before comparing event designs. The model's predictions still require validation.

Common misconceptions and mistakes

Do not replace a distribution of types with an average type without checking the effect on strategic behavior. Best responses need not average in that way.

Do not equate incomplete information with imperfect information. The former concerns missing knowledge about aspects such as payoffs or types; the latter concerns what players observe about actions or histories. A game can involve both.

Do not assume a common prior means identical information. Players can start from the same joint distribution and condition on different observations.

Do not infer equilibrium merely from a plausible strategy description. Incentives must be checked. Likewise, the assumptions of rationality, known rules, and specified beliefs may be demanding in practical settings and should be assessed rather than treated as observations.

Business use

The framework helps businesses state who knows what, when they know it, and how knowledge affects incentives. Those questions matter when designing auctions, contracts, participation rules, and interactions among automated decision makers.

Useful analysis separates the game specification from the proposed behavioral solution. Document the type distribution, correlations, payoff assumptions, and available information, then examine alternative specifications. Empirical validation remains necessary because real participants may learn, misunderstand rules, or have objectives absent from the model.

Check your understanding

Why does the low-cost firm assign its rival an 80% low-cost probability?

Answer: Its own type provides information through the correlated joint distribution.

What invalidates the proposed low-only entry equilibrium?

Answer: A high-cost type gains 0.4 by entering against the proposed rival strategy, instead of receiving zero by staying out.

Related terms

Bayes-Nash equilibrium
Imperfect information
Mechanism design
Expected utility
Private information
type space
common prior
incomplete information

Learn next

Build a small type-probability table and calculate beliefs conditional on each player's information. Then add payoffs, define complete type-dependent strategies, and verify candidate best responses rather than stopping at the model specification.

Sources

Daniel Quint, University of Wisconsin, Lecture 4: Bayesian Games and No Trade, specification of games with incomplete information and conditional type beliefs: https://users.ssc.wisc.edu/~dquint/econ698/lecture%204.pdf The correlated-cost game above is an original teaching example; its failed candidate is intentionally distinguished from an equilibrium.

Previous: Bayesian filtering Next: Bayesian hidden Markov model