Can a losing state have a correct minimax value?

Yes. The value can correctly describe a forced loss; optimal play cannot always avoid losing.

Does a value identify a unique move?

No. Several actions may attain the same value.

Can an opponent's mistake produce an outcome above the value?

Yes, when MAX follows its appropriate strategy in the classical model, weaker opposition can allow a better outcome.

AI Glossary / Graphs and symbolic AI / Intermediate

Minimax value

Definition

The minimax value of a game state is the utility obtained from that state when both players make optimal opposing choices in the specified game model. Using MAX's utility perspective, it is the best outcome MAX can secure against MIN. Terminal states take their assigned utilities; internal states inherit maximum or minimum child values according to who moves.

What it means

A value summarizes an entire continuation tree with one number. It says how the position stands under optimal play, not how attractive its immediate appearance is. The number must be interpreted using the game's utility scale: a value of one could mean a win, one point, or another explicitly defined outcome.

The value belongs to a state under a particular model. Rules, player to move, terminal utilities, and information assumptions all matter. The same visible board arrangement can have a different value when a different player moves next. A heuristic search score may estimate a minimax value without being the exact value of the full game.

How it works

For a terminal state s, set V(s)=U(s), where U is the terminal utility for MAX. For a nonterminal MAX state, set V(s) to the maximum successor value. For a MIN state, set it to the minimum successor value.

These recursive rules assume the tree ultimately reaches evaluated leaves. In the classical full-game interpretation, those leaves are true terminal outcomes. In depth-limited search, an evaluation function supplies estimates at unfinished leaves, and the resulting root score is relative to that cutoff model.

The value is distinct from the action that attains it. Several actions can share one value. Alpha-beta search may also produce bounds for some internal nodes it prunes; an upper or lower bound should not be mistaken for an exact value merely because both are represented numerically.

Three practical examples

Position assessment: label a solved small-game state as a forced win, draw, or loss under an agreed utility scale.
Search explanation: report the value supporting a selected move while identifying whether it is exact, estimated, or bounded.
Regression testing: verify that a search implementation returns the same root value after a harmless change in move order.

Think of it this way

Think of the value as the strength of a promise backed by a complete plan against an opponent's permitted replies. The number represents what can be secured, rather than the most appealing outcome visible somewhere in the possibilities.

The analogy needs qualification: the promise is conditional on the game model and correct strategy execution. It is not a guarantee about an incompletely modeled real-world situation.

Worked example

MAX has two actions from state S. Action A leads to a MIN node with terminal utilities 6 and 2. Action B leads to a MIN node with terminal utilities 3 and 5. All values are measured for MAX.

The child values are V(A)=2 and V(B)=3. Therefore V(S)=max(2,3)=3. MAX can secure 3 by selecting B, and MIN can prevent a higher outcome by choosing its terminal response valued 3. If MIN instead selects 5, the observed outcome is better for MAX, but the original state value remains 3 under optimal opposing play.

Now transform every terminal utility using U'=2U+10. The four utilities become 22,14,16,20. The child values become 14 and 16, and the root value becomes 16, exactly 2×3+10. The preferred action remains B. The numerical value changes with the scale even though this positive affine transformation preserves preferences.

Finally, suppose the terminal numbers were only heuristic evaluations of unfinished positions. The same backups would still give 3 on the original scale, but that would be the depth-limited search value. Without further evidence, it would not establish the exact result achievable in the complete game. Arithmetic correctness and evaluation accuracy are separate issues.

Illustrative case study

Fictional case: an educational game displays a score of “+3” beside a suggested move. Students interpret it as a guaranteed three-point victory, although the score comes from an early search cutoff and a position-evaluation formula.

The designers revise the display to distinguish estimated advantage from solved terminal value. They retain the score's perspective and search depth in the explanation and reserve “forced win” for positions supported by the appropriate analysis. The fictional case illustrates clearer interpretation, not a measured learning outcome.

Common misconceptions and mistakes

A minimax value is not necessarily an expected value. In a deterministic adversarial tree, minimum and maximum choices replace probability-weighted averaging. Chance nodes require a model that accounts for their probabilities.

Nor is a positive score automatically a win. That interpretation depends on the utility scale and whether the value is exact. A large heuristic advantage can still overlook a decisive continuation beyond the horizon.

Keep the perspective consistent. A score reported for MAX cannot be interpreted as MIN's utility without the appropriate relationship between their payoffs. The classical scalar formulation assumes directly opposing preferences; general-sum games may require separate utilities for different players.

Business use

For game products and adversarial simulations, value reporting should explain the score scale and level of certainty. An exact solved value, a heuristic estimate, and a pruning bound serve different purposes in a decision interface.

When comparing versions of an agent, confirm that the utility scale and evaluation conventions are unchanged. A higher displayed number after rescaling is not evidence of better decisions. Compare actual choices and outcomes using consistent rules and relevant evaluation cases.

Check your understanding

A MAX node has exact child values −5 and −1. What is its value, and does that necessarily mean victory?

Answer: Its value is −1. Whether that represents victory depends on the defined utility scale; it is simply the better of those two values.

Related terms

Minimax decision
Minimax search
Utility
Terminal states
Evaluation function
Zero-sum games
terminal state
game value
alpha-beta bound
zero-sum game

Learn next

compute values on a tiny game tree using win/draw/loss utilities. Then replace terminal outcomes with heuristic cutoff scores and explain precisely which guarantee has been lost.

Game tree

Sources

University of California, Berkeley, CS 188, develops terminal utilities and recursive minimax state values: https://inst.eecs.berkeley.edu/~cs188/textbook/games/minimax.html

Poole and Mackworth explain adversarial game values, pruning bounds, and estimated leaf evaluations: https://artint.info/3e/html/ArtInt3e.Ch14.S3.html

The utility transformation and fictional display case are original teaching examples.

Previous: Minimax search Next: Minimization (clinical trials)