Consider. Compare. Choose.

Rules, search and evaluation

A chess program needs three different abilities. It must know which moves are legal, explore what can happen after those moves, and compare positions it has not fully solved. Improving one ability does not remove the need for the others.

An evaluation is a judgement in numerical form. A simple one might value a queen more highly than a pawn. A richer one might also recognise an exposed king or an inactive rook. The number is useful only in the context of the system that produces it.

Minimax: plan for resistance

Imagine evaluating from White’s perspective. White prefers larger values; Black prefers smaller ones. Working backwards through a tree of continuations alternates these choices. That procedure is minimax. It assumes the opponent selects a strong reply rather than co-operating with your plan.

Alpha-beta: less work, the same answer

Alpha-beta pruning keeps bounds on what a branch can achieve. If a continuation cannot improve the decision already available, the program can stop examining it. For the same evaluation and search depth, correct alpha-beta pruning preserves the minimax result. Searching promising moves first often makes it more effective.

Why stop in a quiet position?

A fixed-depth search can stop halfway through an exchange. Counting material after taking a rook, but before the opponent recaptures a queen, would be misleading. Tactical extensions continue selected forcing sequences before settling on a value.

Our reconstruction follows captures judged significant and responds to checks beyond its ordinary two-ply search. It also has explicit limits. The reconstruction notes describe those choices.

From explicit rules to learned evaluation

Shannon’s 1950 paper set out a systematic discussion of computer chess, including search and evaluation. Deep Blue later combined substantial computing power with specialised chess engineering.

Stockfish’s introduction of NNUE in 2020 illustrates a different development: a learned evaluation used within an alpha-beta search. “Neural” does not mean search has disappeared.

AlphaZero combined learning through self-play with neural-network guidance and a different search approach, Monte Carlo tree search. It was given game rules; its training did not begin with a hand-written list of chess positional preferences.

No single ladder to intelligence

These approaches answer different engineering questions. They differ in training cost, hardware, search procedure and the knowledge built into them. A useful comparison asks how a system selects a move and what evidence supports its judgement, rather than treating every later program as the same design made faster.