Value network in chess: neural evaluation for positions

Value Network

Definition

A value network is a machine-learned evaluation function that estimates the expected game outcome (win, draw, or loss) from a given chess position, usually from the perspective of the side to move. Its output is commonly a number in the range [-1, 1] (loss to win) or [0, 1] (0% to 100% win probability), whereas traditional engines report centipawns (cp). Unlike hand-crafted evaluations that sum material and positional terms, a value network learns patterns directly from data.

Modern systems often pair a value network with a policy network (which suggests promising moves), or use a neural evaluation embedded in an alpha–beta engine (as in NNUE).

How It’s Used in Chess

• Guiding search:
  • Monte Carlo Tree Search (MCTS): Engines like Leela Chess Zero evaluate leaf positions with the value network and back up results through the tree, focusing exploration on lines with higher expected score. See Monte Carlo Tree Search.
  • Alpha–beta search: Engines with NNUE-style nets (e.g., Stockfish since 2020) use a neural evaluation at leaves and throughout the search to inform pruning and move ordering.
• Move ordering and pruning: A reliable value estimate helps prioritize critical candidate moves and trim unpromising branches earlier, enabling deeper, more relevant analysis.
• Training signal: Value networks are trained on positions paired with eventual game results from self-play or databases. The network learns to predict the result z from a position s by minimizing error between its prediction V(s) and z.
• Interpreting scores: Value networks output win/draw/loss probabilities; GUIs may map these to centipawns using engine-specific conversions (often logistic-shaped). The mapping is not universal.

Strategic and Historical Significance

• Learned evaluation breakthrough: Value networks marked a shift from hand-tuned heuristics toward learned, pattern-rich assessments of king safety, initiative, compensation, and long-term plans.
• AlphaZero era: In 2017, AlphaZero’s policy–value network plus MCTS defeated top engines. Its value network favored dynamic, long-term compensation and pressure that classical terms often underweighted.
• Community impact: Leela Chess Zero (2018–) democratized policy/value training via distributed self-play. Stockfish’s NNUE integration (2020) showed neural evaluation also excels within alpha–beta frameworks.
• Influence on style: Engine games guided by value networks popularized exchange sacs, long squeezes, and strategic pawn breaks that emphasize initiative over short-term material.

Examples

Example A — Compensation over material (Benko Gambit). Black sacrifices a queenside pawn for lasting activity and pressure. A value network often rates Black’s practical chances highly even before concrete tactics appear:

Idea: Black gains open a- and b-files, rapid development, and pressure on the queenside dark squares. Although down a pawn, long-term activity can translate to a strong winning expectancy.

Example B — Endgame mastery (Lucena-type R+P vs R). Tablebases confirm these positions as theoretical wins; value networks learn to score them near a full point for the stronger side:

Idea: With the pawn on the seventh rank and the king well placed (here on c6), White can “build a bridge” with the rook to escape checks and promote. A value network, trained on many such endings (and often informed by tablebases), recognizes the decisive advantage.

Interesting Facts and Anecdotes

• AlphaZero vs. Stockfish, 2017: AlphaZero’s value network frequently preferred quiet buildup and piece activity over immediate material gains, culminating in powerful attacks and endgame conversions.
• Calibration caveat: A “0.80” value is a win probability, not “0.80 pawns.” Conversions to centipawns are engine- and UI-specific.
• Edge cases: Fortresses, stalemate tricks, and 50-move-rule nuances can challenge learned evaluations. Projects mitigate this with curated data and tablebase integration.

Related Concepts

• policy network — proposes moves to explore during search.
• Monte Carlo Tree Search — search guided by policy and value estimates.
• NNUE — efficiently updatable neural evaluation within alpha–beta engines.
• evaluation — the broader notion of scoring positions (centipawns, probabilities, etc.).

Practical Notes

• Perspective: Values are almost always from the side to move. A high value on White’s turn flips if Black gets the move in the same position.
• Heuristic vs. proof: Value networks highlight promising directions (space, king safety, initiative). In sharp tactics, the search still verifies concrete lines.
• GUI interpretation: If your engine shows only cp, remember that a displayed +1.50 may be a mapped or combined signal, not the raw value-network output.