Minimax assumes the opponent is rational, adversarial, and plays optimally to minimize your outcome. That means at every opponent turn, Minimax behaves as if the opponent will choose the move that is worst for you, not the move that is “likely,” “human-like,” or “average.” This is a worst-case model: if your move still looks good under the opponent’s best defense, it’s considered a good move. This assumption is often exactly what you want in competitive games or security-style problems, but it can be overly conservative in casual settings.
This “optimal opponent” assumption implies several practical constraints. First, the game should be close to zero-sum or at least strongly competitive; otherwise “minimizing your utility” may not reflect the opponent’s true goal. Second, Minimax does not account for opponent mistakes unless you explicitly model them (for example, by using a probabilistic opponent rather than a hard minimizer). Third, Minimax assumes the opponent has the same information about the game state as you do (perfect information). If the opponent’s knowledge differs (hidden cards, fog of war), the assumption “they pick the worst response” becomes ambiguous because “worst response” depends on what they know and believe. Finally, the assumption can create “safe” play: Minimax may prefer a guaranteed draw over a risky line that often wins against humans but loses against perfect defense.
A simple scenario shows the effect. Imagine you can choose Move A which guarantees a draw (utility 0), or Move B which wins if the opponent blunders (+10) but loses if they defend correctly (-10). Minimax will choose Move A because it maximizes the guaranteed minimum outcome (0 is better than -10). That’s correct under the assumption, but it may feel “too cautious” against real people. If you want more human-like aggression, you’re changing the opponent model, not “improving” Minimax. Outside games, the same assumption can represent worst-case uncertainty: you’re not fighting a person, you’re defending against the most harmful plausible scenario (bad data, ambiguity, edge cases). In retrieval pipelines using Milvus or Zilliz Cloud, a Minimax-like opponent can represent “the retrieved snippet that looks relevant but is actually misleading,” pushing you toward decisions that remain acceptable even if the top hit is wrong. That’s useful when you care more about avoiding catastrophic errors than maximizing average relevance.