How do chess engines evaluate a position?

By combining a search through possible future moves with an evaluation of the resulting positions — and the way that evaluation is produced changed fundamentally in recent years.

The classical approach, which dominated for decades:

Search. Examine the tree of possible moves — my move, your reply, my response — to a certain depth. The tree grows explosively, so the key technique is alpha-beta pruning: once a line is proven worse than one already examined, it can be abandoned without further analysis. This eliminates most of the tree without affecting the result.

Evaluation. At the end of each line, score the position using hand-written rules: material value of pieces, pawn structure, king safety, piece activity, control of the centre, and dozens of other heuristics weighted by human experts.

Quiescence search extends analysis through sequences of captures, to avoid stopping in the middle of an exchange and misjudging the position badly.

Opening books and endgame tablebases supply perfect play at both ends — tablebases contain solved results for positions with few pieces, and are genuinely complete rather than approximate.

Deep Blue beat Kasparov in 1997 using essentially this approach at enormous scale.

The neural network approach. AlphaZero, in 2017, learned chess entirely from self-play with no human knowledge beyond the rules. Instead of hand-written evaluation, a neural network learned to judge positions, and it searched far fewer positions per second than classical engines while playing more strongly — evaluating better rather than calculating more.

Its play was notable for accepting long-term positional compensation that classical engines undervalued, and for unusual pawn sacrifices.

Where things stand now. The strongest engines combine both: Stockfish adopted NNUE, an efficiently updatable neural network evaluation, running on conventional hardware alongside its traditional search. The synthesis outperforms either approach alone.

What the numbers mean. Evaluation is expressed in pawn equivalents — +1.5 means an advantage worth about a pawn and a half. Engines also report depth in plies (half-moves) and the principal variation, their expected best line.

Engines have changed how the game is studied, and human opening preparation now largely follows them.