Can a game have perfect information and uncertain outcomes?

Yes. Future chance events and future player choices can remain uncertain.

Is poker a perfect-information game?

Standard poker involves hidden cards, so it is an imperfect-information game.

Does perfect information imply perfect play?

No. A player may observe the history and still select poor actions.

AI Glossary / Graphs and symbolic AI / Beginner

Perfect information

Definition

Perfect information is a property of a sequential game in which the player making a decision knows the history of play up to that point. In an extensive-form representation, each decision information set contains just one node.

What it means

A player can distinguish exactly which decision situation has been reached. In chess, the board and previous moves are visible; an opponent's hidden cards in poker create a different kind of information structure.

Perfect information does not mean that a player knows future choices or can calculate their consequences easily. It also does not exclude chance: a publicly observed die roll can be part of a perfect-information game, even though its result was uncertain before the roll.

Complete information is a separate concept concerning knowledge of the game's structure, payoffs and relevant player characteristics. Participants can know the rules and payoffs while remaining unable to observe another player's action. The present entry uses the game-theoretic sense, rather than the broader everyday phrase meaning all useful facts.

How it works

Represent the game as a sequence of decisions and, where relevant, chance events. Each nonterminal history identifies whose turn it is and what actions are available.

For each decision, specify what the acting player has observed. If the player can distinguish the current node from every alternative node, its information set is a singleton. If two or more possible histories are indistinguishable at that decision, the game has imperfect information there.

This distinction constrains strategies. A player cannot select different actions at two nodes it cannot tell apart simply because an analyst can see both nodes in a diagram. Treating hidden information as visible gives the modeled player an advantage it does not possess.

Suitable solution methods depend on additional assumptions, such as whether the game is finite, zero-sum or stochastic. Perfect information alone does not specify a universal algorithm or make large games easy.

Three practical examples

1. Board-game agents: A chess program receives the visible game position and relevant history. Its difficulty concerns selecting a strong move, rather than estimating hidden opponent cards.

2. Sequential allocation: Two teams choose resources one after another, with the second team observing the first selection. The model differs from simultaneous sealed selections even when the choices and payoffs are identical.

3. Simulation design: A training environment reveals a random event to every player before the next action. Chance remains present, but that revealed event does not create hidden information at the subsequent decision.

Think of it this way

Think of players sharing a complete recording of everything that has happened on the game board. They can still disagree about what will happen next and how best to respond.

The analogy concerns access to history, not unlimited reasoning power. Having the recording does not reveal private future plans, guarantee foresight or make one player capable of winning from every position.

Worked example

Consider a two-stage toy game. Player A chooses Left or Right. Player B then chooses Match Left or Match Right and receives one point if the choice matches A's earlier action, otherwise zero.

If B observes A's action before choosing, B can match it in either case and receives one point. The two decision nodes are distinguishable, so B has a separate choice for each observed history.

If A's action is hidden, B cannot condition its move on which node was reached. Suppose A independently chooses Left and Right with equal probability. Either fixed guess by B has expected payoff 0.5 × 1 + 0.5 × 0 = 0.5. Randomizing the guess cannot exceed 0.5 against that specified distribution.

The difference comes from observation, not a change in the available actions. The hidden-action version must represent B's two nodes in one information set.

Illustrative case study

In a fictional project, Maple Games evaluates an agent in a card-game simulator. The internal state includes every player's cards, and an early observation interface accidentally passes that full state to the agent.

The team separates the simulator's complete state from the observations allowed by the game. It verifies that the policy receives only permitted information and retests decisions that previously depended on hidden cards.

The case illustrates an evaluation error: a capable policy under perfect observation may not be a valid policy for the intended imperfect-information game. Its earlier results cannot establish performance under the corrected information constraints.

Common misconceptions and mistakes

Perfect information means knowing future moves. It concerns the observed history at a decision.
Chance makes information imperfect. Publicly observed chance outcomes can coexist with perfect information.
Knowing the rules is enough. Players may still be unable to observe past actions or private outcomes.
Every node drawn separately is distinguishable to the player. Information sets specify what the player can actually tell apart.
A perfect-information game must be easy to solve. Its decision tree can remain enormous.

Business use

Information structure matters in strategic simulations, multi-agent systems and workflow experiments. A model can produce misleading conclusions if one participant receives information unavailable in the real process.

Document who observes each action, when it becomes visible and which random outcomes remain private. Delayed reports can matter as much as permanently hidden facts. Keep the simulator's internal state separate from each agent's observation interface.

When comparing strategies, hold those information permissions consistent. A result obtained with additional knowledge may be useful as an upper benchmark, but it should not be presented as achievable performance for an agent that cannot access that knowledge.

Check your understanding

Why can B not simply use the successful matching strategy when A's action is hidden?

Answer: That strategy requires choosing differently after Left and Right. B cannot distinguish those histories, so the strategy is not implementable with B's available information.

Related terms

Information sets
Imperfect information
Strategy
Minimax
Chance nodes
Information set
extensive-form game
complete information
chance node
partial observability

Learn next

Draw a two-stage decision tree and mark which earlier events the second player observes. Merge indistinguishable decision nodes into information sets, then identify a strategy that becomes invalid when an observation is removed.

Sources

Russell and Norvig distinguish perfect-information, chance and imperfect-information games and discuss associated search methods: https://aima.cs.berkeley.edu/slides-pdf/chapter06.pdf

The matching game and Maple Games case are original illustrations. The numerical comparison assumes the explicitly stated equal-probability hidden choice.

Previous: Perfect hash function Next: Perfect rationality