Can two different actions have the same value?

Yes. They can produce different distributions of outcomes with the same backed-up value.

Can a value be negative?

Yes. The sign depends on the utility scale and the outcomes available from that position.

Does the value include the opponent's choices?

Yes. MIN backups represent the opponent selecting the lowest successor value for MAX under the classical model.

AI Glossary / Graphs and symbolic AI / Intermediate

Expectiminimax value

Definition

An expectiminimax value is the value assigned to a game-tree position by combining maximizing choices, minimizing choices, and probability-weighted chance outcomes. It expresses the expected utility available to the maximizing player under the specified game model and adversarial responses. It is the numerical result of expectiminimax evaluation, rather than the search procedure itself.

What it means

The value of a position depends on what can happen next and who controls each transition. If the maximizing player chooses, the value is the best successor value. If the minimizing opponent chooses, it is the worst successor value from the maximizing player's perspective. If chance determines the transition, it is the weighted average.

Terminal positions receive their stated utility directly. Values at earlier positions summarize the remaining game under those backup rules.

An exact value is conditional on the rules, probabilities, payoff scale, and information available to each player. It does not mean the maximizing player will receive that exact number in every play. A depth-limited calculation may instead produce an estimated expectiminimax value based on heuristic leaf evaluations.

How it works

For a terminal node s, set its value to the terminal utility U(s). For a MAX node, take the maximum of its successor values. For a MIN node, take their minimum.

For a chance node, sum P(t | s) times the value of successor t across all possible outcomes. These probabilities belong to the random event at that node; they are not preferences over the outcomes.

Work upward until the desired position has a value. A root action can then be chosen by comparing its resulting successor values. Ties can leave several equally valued choices under the model.

Because expectations use magnitudes, the numerical utility scale matters. A positive affine transformation preserves rankings, but a general nonlinear transformation can change them. A correct value also depends on preserving decision timing: choosing after observing a random event can provide more options than choosing before observing it.

Three practical examples

1. A game agent attaches values to candidate positions so that it can compare moves involving both dice outcomes and opponent responses.

2. A search debugger compares computed node values with a small hand-worked tree to identify an incorrect probability or backup operation.

3. A game-analysis tool distinguishes an exact solved-tree value from an estimate obtained at a search cutoff, making the interpretation of its displayed score clearer.

Think of it this way

Think of a position's value as a summary written at the top of a branching decision sheet. The summary reflects the choices and uncertainties below it, rather than simply copying the largest visible prize.

The analogy is limited because changing who chooses a branch or when information becomes available changes the summary. A value without its underlying game assumptions is not a complete description of a position.

Worked example

Consider a position where MIN chooses either to end the game with payoff 6 to MAX or to continue to a small subgame. In that subgame, MAX must choose L or R before an equally likely random outcome W1 or W2 is revealed.

Under L, the payoff is 10 for W1 and 0 for W2, so L has value 5. Under R, the payoffs are 4 and 6, also giving value 5. MAX's subgame value is therefore max(5, 5) = 5. At the original position, MIN chooses min(6, 5) = 5.

Now change the rules so that W1 or W2 is revealed before MAX chooses. For W1, MAX chooses L and gets 10. For W2, MAX chooses R and gets 6. The subgame value becomes 0.5 × 10 + 0.5 × 6 = 8. MIN can still end the game at 6, so the original position's value becomes min(6, 8) = 6.

The same payoff table yields different values because the information timing changed. Maximizing an expectation and averaging state-specific maxima are different operations. These are exact values for the toy trees.

Illustrative case study

Fictional case: Cedar Games displays a score beside each suggested move. During testing, two analysts obtain different values from the same payoff table. One lets the player choose after seeing a random event; the other requires an earlier commitment.

They draw the decision order explicitly and label which nodes represent observed outcomes. The discrepancy disappears once both evaluate the same game. The case illustrates that a numerical disagreement can originate in the modeled information structure rather than in arithmetic or search code.

Common misconceptions and mistakes

Do not confuse a position's value with the payoff of its most likely continuation. A chance-node value averages all modeled outcomes.

Do not assume an exact expectation guarantees that amount on each play. The realized payoff can differ substantially.

Avoid comparing scores from different utility scales as though they were interchangeable. Even identical-looking numbers can represent different objectives.

Do not present a heuristic cutoff value as a proved game-theoretic value. The backup may be correct while the leaves remain uncertain estimates. Also, hidden information cannot generally be represented by simply allowing independent choices in states the player cannot distinguish.

Business use

Expectiminimax values are useful in game AI, educational tools, and deliberately adversarial simulations. They provide interpretable summaries when the model's assumptions are visible.

When displaying a value, identify its payoff convention, search scope, and whether it is exact or estimated. Test sensitivity to event probabilities and opponent assumptions. In a business simulation, first establish whether treating another actor as a minimizing opponent is appropriate; the value otherwise answers a question that may not match actual behavior.

Check your understanding

Why is the first subgame's value 5 rather than 8 when MAX must choose before seeing W1 or W2?

Answer: MAX must commit to one action for both possible outcomes. The value 8 incorrectly allows a different action after each outcome is revealed.

Related terms

Learn next

Practice calculating terminal, MAX, MIN, and chance values separately. Then change a tree's decision order while retaining its payoff table. Explain which information each player has at each choice and why the resulting values can change.

Sources

Primary source: Russell and Norvig, Game Playing, expectiminimax and the importance of evaluation magnitudes: https://aima.cs.berkeley.edu/slides-pdf/chapter06.pdf

The timing comparison and fictional case are original teaching examples. The entry concerns the value produced by the classical game-tree model, with approximate evaluations explicitly distinguished.

Previous: Expectiminimax Next: Experience replay