Probability matching¶
A decision strategy that randomizes predictions or choices in proportion to estimated outcome probabilities rather than always choosing the most likely option.
Core Idea¶
Probability matching converts a probability distribution into a randomized decision distribution. If outcomes occur with probabilities 0.6 and 0.4, a matcher chooses the corresponding labels with those same probabilities. This differs from maximizing expected zero–one accuracy, which chooses the 0.6 label every time when no other information exists.
The strategy's quality depends on the objective and information structure. In the simple independent binary case, matching succeeds with p²+(1−p)² and is generally dominated by taking the modal class. Yet matching remains important in behavioral research and as a contrast for sequential randomized policies. Thompson sampling is related but samples from uncertainty about which action is optimal and updates through feedback.
Scope of Application¶
- Behavioral experiments. Human prediction frequencies are compared with observed base rates.
- Classification. Randomized labels are generated from calibrated class probabilities.
- Decision theory. Matching and maximizing are contrasted under explicit loss.
- Sequential choice. Related posterior-sampling policies are distinguished from simple matching.
Clarity¶
State alternatives, probability source, conditioning information, update rule, randomization independence, trial horizon, payoff or loss, and whether matching is assessed in expectation or realized frequency. Keep probability calibration distinct from decision optimality. Inclusion test: A strategy probability-matches when its long-run choice distribution intentionally follows the estimated outcome distribution rather than only its rank ordering. Exclusion test: A deterministic maximum-probability rule is excluded even if its overall success rate happens to equal a probability. Nearest boundary: Thompson sampling is nearby because it samples a posterior over optimal actions, but it incorporates sequential learning and action value rather than merely matching outcome base rates. Exit condition: The identity exits when choice proportions are unrelated to probabilities, randomization is accidental, or the strategy always takes the argmax. Common misclassifications: It is not probability estimation alone. It is not always choosing the most probable class. It is not automatically irrational under every payoff or dependent environment. It is not synonymous with Thompson sampling. Nearest named distinctions: Probability calibration: Compares predicted probabilities with outcome frequencies; it does not require randomized actions. Expected-utility maximization: Selects the action with greatest expected value under the utility model. Thompson sampling: Samples by posterior probability of action optimality while learning sequentially. Random guessing: May ignore estimated base rates entirely.
Manages Complexity¶
One rule compresses an uncertain distribution into stochastic behavior that preserves aggregate proportions. The compression hides order effects, dependence, learning, individual heterogeneity, utility, and sampling noise; those features determine whether the behavior is descriptive, normative, or strategically useful.
Abstract Reasoning¶
- Define mutually exclusive outcomes and the available information.
- Estimate normalized base or conditional probabilities.
- Specify a randomization device using those probabilities.
- Generate choices without replacing matching by argmax.
- Calculate expected payoff under the actual loss and dependence assumptions.
- Compare with maximization and alternative adaptive policies.
- Check realized frequencies and update only if the strategy includes learning.
Knowledge Transfer¶
The probability-to-action mapping transfers to any finite choice set with meaningful probability estimates. Claims of optimality do not transfer across loss functions, temporal dependence, competition, or learning environments. The cargo is proportional randomized choice, not the mere use of probabilities.
Neighborhood in Abstraction Space¶
Probability matching sits in a crowded region of the domain-specific corpus (19th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.
Family — Decision & System Modeling Frameworks (30 abstractions)
Nearest neighbors
- Two-Moment Decision Model — 0.91
- ÉLECTRE — 0.91
- Fitness-Proportionate Selection — 0.91
- Chance-Constrained Programming — 0.90
- Approximate Bayesian Computation — 0.90
Computed from structural-signature embeddings · 2026-10-08