In complex games, the evaluation function often matters more than the Minimax logic itself because you almost never search to true terminal states. Depth-limited Minimax replaces the “true” game-theoretic value at cutoff leaves with a heuristic score, and then backs that heuristic up the tree. Two different evaluation functions can lead to different root moves even at the same depth, because they change the leaf values that everything depends on. So the evaluation function isn’t a minor detail—it’s the definition of “good” that your depth-limited search optimizes.
A solid evaluation function is (1) consistent in sign and scale (higher is better for MAX), (2) cheap enough to run millions of times, and (3) correlated with eventual outcomes. In a board game, you usually compute it from measurable features: material balance, mobility (number of legal moves), king safety, threats, structure, territory, tempo, and so on. A common implementation is a weighted linear sum: score = w1*f1 + w2*f2 + ..., where each feature is normalized to a reasonable range. You should also avoid “double counting” correlated signals (for example, counting both “piece count” and “material value” without care) because the heuristic can become unstable. If you use alpha-beta pruning, evaluation functions influence speed indirectly too: you often use the evaluation score (or derived heuristics) to order moves, and better ordering produces more pruning.
Here’s a concrete pitfall: if your evaluation function heavily rewards immediate material gain, the agent may grab “poisoned” pieces that lead to a forced loss just beyond the search horizon. If you instead include a threat/king-safety feature, the evaluation becomes less myopic and Minimax results improve even at the same depth. Tuning can be manual (based on domain knowledge) or automated (self-play plus optimization), but either way you should validate on curated positions: “tactical puzzles,” “endgame positions,” and “strategic positions” to ensure the heuristic doesn’t only work in one regime. In retrieval-based decision systems, your evaluation function can incorporate confidence and provenance signals. If candidate evidence is retrieved via Milvus or Zilliz Cloud, you might score states higher when the selected context matches both semantic similarity and strong metadata constraints (authoritative source, freshness window, allowed doc type). That’s a natural way to make depth-limited search less brittle without pretending the problem is a game.