Is type A the same as breadth-first search?

No. The tree can be processed depth first while still considering every variation to the same horizon.

Does increasing depth always solve evaluation problems?

No. It changes the boundary but can still leave important consequences beyond it, and the computation grows rapidly.

How does type B differ?

Shannon's type B approach selects variations and extends important lines toward suitable evaluation points rather than using only uniform bounded coverage.

AI Glossary / Graphs and symbolic AI / Intermediate

Type A strategy

Definition

A type A strategy is Claude Shannon's historical term for a game-playing approach that considers all move variations to a predetermined depth and then chooses a move using backed-up position evaluations. It is associated with full-width, depth-limited minimax search.

What it means

The strategy sets a search horizon and gives every legal continuation within that horizon a place in the decision calculation. At the boundary, an evaluation function estimates how favorable the position is. These estimates are combined under the assumption that the opponent chooses replies that are worst for the player.

The term comes from Shannon's 1950 paper on computer chess. It does not refer to a personality type, a business strategy classification or a machine-learning model family.

Considering all variations to a limit is different from solving the entire game. A depth-limited result depends on the quality of its boundary evaluations and can miss consequences beyond the horizon. Its systematic coverage is useful, but the limit remains a real limitation.

How it works

Generate legal moves from the current position, then generate legal responses and continue until the chosen depth or an actual terminal state is reached. Evaluate terminal results appropriately and use a heuristic score for nonterminal boundary positions.

Back up scores through the tree. At the player's decision points choose the maximum; at the opponent's points choose the minimum when scores are expressed from the player's perspective. The selected root move has the best backed-up value under those assumptions.

A simple implementation examines the full tree to the limit. Sound alpha-beta pruning can avoid branches that provably cannot change that same minimax result; this differs from excluding a move merely because it looks unpromising. The historical type A idea concerns systematic bounded lookahead, while modern implementations may combine many additional techniques.

Three practical examples

A board-game teaching program uses a fixed-depth minimax search to show how an opponent's replies affect move selection.
An engine developer uses a full-width bounded search as a comparison when testing a selective search policy.
An AI workshop measures how increasing the search depth expands the number of positions requiring evaluation in a small game.

Think of it this way

Think of reviewing every branch of a decision diagram up to a marked line, then estimating the outcomes beyond that line. You do not skip a branch just because the first step looks unattractive.

The analogy is limited to adversarial decisions with the relevant minimax assumptions. Ordinary business options are not always opposed by an actor trying to minimize your score, and a depth boundary does not make uncertain future consequences disappear.

Worked example

Backing up all responses at a fixed horizon

At the root, player MAX can choose A, B or C. Each move lets opponent MIN choose one of three replies. Suppose the resulting boundary scores, all from MAX's perspective, are:

A: 7, 2, 5.
B: 4, 3, 6.
C: 9, −1, 8.

MIN's best response gives each root move its minimum: A becomes 2, B becomes 3 and C becomes −1. MAX therefore chooses B, whose worst represented reply is best.

There are nine evaluated leaves in this two-ply example, and thirteen nodes including the root and its three children. Choosing C because it contains a leaf scored nine would ignore the opponent's ability to select −1.

These are invented boundary estimates. Unless the leaves are genuine final outcomes, the calculation proves only which move maximizes this bounded minimax evaluation, not which move wins the full game.

Illustrative case study

Fictional Alder Games builds a small strategy-game opponent using fixed-depth search. Testers find that it sometimes chooses an immediate gain followed by a loss just beyond the search boundary.

The team keeps the baseline calculation intact for comparison, then studies whether deeper searches or a better evaluation function address those positions. It records search depth and evaluation settings alongside results, rather than describing the original algorithm as having considered every possible consequence.

The fictional case illustrates why complete coverage within a horizon can coexist with a poor decision. It contains no measured playing-strength improvement and does not claim that a particular depth is sufficient for all positions.

Common misconceptions and mistakes

Type A means searching to the end of every game. The historical strategy uses a definite lookahead limit.
Full-width search needs no domain knowledge. A practical boundary evaluation still needs a meaningful way to score positions.
Any skipped branch makes search selective in the risky sense. Alpha-beta pruning can preserve the fixed-depth minimax result when applied correctly.

Business use

The concept gives teams a clear baseline for discussing search coverage. It helps distinguish a method that systematically handles all bounded alternatives from one that depends on a move-selection heuristic.

For a game or simulation product, report depth, node counts, time limits and evaluation assumptions. Compare decision quality on positions that expose delayed consequences, not only on routine examples. If a business simulation uses minimax, first justify its adversarial model; the historical chess strategy does not automatically fit cooperative or probabilistic environments.

Check your understanding

A root move has reply scores 10, 6 and −4 in a two-player minimax example. What value is backed up at its opponent node?

Answer: −4. The opponent selects the least favorable represented outcome for the root player, so the attractive score of ten does not determine that move's value.

Related terms

Learn next

Read minimax and type B strategy. Calculate all backed-up values in a small two-ply tree, then change one previously overlooked opponent reply and observe how the root choice changes.

Sources

Claude E. Shannon, Programming a Computer for Playing Chess, 1950, Sections 4–6, original paper available through the Computer History Museum: https://www.computerhistory.org/chess/doc-431614f453dde/

Russell and Norvig, Adversarial Search, minimax and alpha-beta discussion: https://aima.cs.berkeley.edu/4th-ed/pdfs/newchap06.pdf

The linked original Shannon PDF and relevant textbook passages were read. The thirteen-node example and fictional case are original; no actual chess-engine result or modern implementation taxonomy is asserted.

Previous: Type-2 fuzzy sets and systems Next: Type B strategy