Heuristics

Heuristics are cognitive shortcuts, rules of thumb, or simplified decision procedures that enable agents to make reasonable judgments under conditions of bounded rationality — limited time, information, and computational resources. The term derives from the Greek heuriskein (to find or discover), the same root as eureka. In cognitive science, heuristics are studied as both the cause of systematic cognitive bias and as the mechanism of remarkable adaptive intelligence. In mathematics and computer science, they are search strategies that find good-enough solutions when finding the optimal solution is computationally infeasible. These two uses are connected by a common insight: in complex systems with high-dimensional search spaces, exact optimization is often impossible, and heuristics represent the pragmatist's answer to intractability.

Two Research Programs in Conflict

The dominant research program in heuristics research — launched by Amos Tversky and Daniel Kahneman in the early 1970s — treats heuristics primarily as sources of systematic error. The heuristics-and-biases program documents the ways in which cognitive shortcuts produce predictable deviations from rational norms: availability bias (judging probability by how easily examples come to mind), representativeness (judging category membership by similarity to stereotypes), and anchoring (insufficient adjustment from initial estimates). The program is empirically rich, methodologically sophisticated, and has produced robust findings across decades.

It has also been systematically misread. The heuristics-and-biases program measures deviation from normative models — usually expected utility theory or Bayesian probability. But this measurement frame presupposes that the normative models are the right standard of comparison. Gerd Gigerenzen and the ABC Research Group have argued forcefully that this presupposition is wrong: in real-world environments, fast and frugal heuristics — decision procedures that use minimal information and computation — often outperform complex Bayesian optimization, because the Bayesian calculation requires accurate probability estimates that are unavailable in real environments, and errors in those estimates compound into worse-than-heuristic outcomes. The ecological rationality of a heuristic is not its match to a formal norm but its fit to the structure of the environment in which it operates.

This is not merely an empirical dispute. It is a dispute about what rationality means in systems embedded in environments — and the pragmatist reading of this dispute is unambiguous: a decision rule that reliably produces good outcomes in its ecological niche is rational for that niche, regardless of whether it satisfies axioms designed for idealized agents in stipulated probability spaces.

Heuristics in Formal Systems

In computer science and operations research, a heuristic algorithm is a problem-solving method designed to find a good-enough solution in a reasonable time when exact methods are computationally infeasible. The distinction matters because many practically important problems — the traveling salesman problem, protein structure prediction, scheduling, combinatorial optimization — are NP-hard: no known algorithm solves them exactly in polynomial time.

Heuristic approaches for such problems include:

• Greedy algorithms: at each step, take the locally optimal choice, accepting that the globally optimal path may not be found.
• Simulated annealing: accept worse solutions with a probability that decreases over time, allowing the search to escape local optima.
• Genetic algorithms: maintain a population of candidate solutions, recombine and mutate them, and select for fitness — a computational implementation of evolutionary heuristics.
• Branch and Bound: explore the solution tree but prune branches that cannot improve on the current best solution.

These methods are successful precisely because they accept the pragmatist constraint: the goal is not the best possible solution but the best solution findable in available time with available resources. The formal computer science concept of a heuristic is therefore not an approximation that falls short of a standard — it is the standard, appropriately stated for the actual problem.

Scientific Heuristics

Scientists use heuristics that are rarely made explicit but are nonetheless constitutive of how science progresses. Occam's Razor (prefer simpler explanations) is a heuristic, not a derivable law. The practice of seeking mechanistic explanations (not merely statistical associations) is a heuristic. The preference for theories that make novel predictions is a heuristic. These are not arbitrary rules of thumb — they are the accumulated procedural knowledge of a community that has learned, through centuries of practice, which search strategies tend to find true theories more reliably than others.

This is the systems insight that the heuristics literature has not fully absorbed: heuristics are not deviations from optimal reasoning. They are the evolved or designed structure of a cognitive or algorithmic system for navigating a particular search space. Understanding why a heuristic works — what environmental structure it exploits, what statistical regularities it relies on — is the science of adaptive cognition. Understanding when it fails — which environmental structures violate its assumptions — is the map of its limits.

A theory of heuristics that only catalogs failures without explaining why the heuristics work at all is not a theory. It is a list of complaints against a form of intelligence that has survived because it works. Any model of cognition that treats every deviation from Bayesian rationality as an error rather than as information about the structure of the cognizer and its environment has confused the map for the territory — and that confusion is itself a bias the literature has not corrected.