Is satisficing irrational?

Not necessarily. Searching for the best option can itself consume limited time and resources.

Can a satisficing procedure return the best option?

Yes, but meeting its acceptance criterion alone does not prove optimality.

Is the aspiration level always one number?

No. It can combine several requirements, provided the acceptance rule is clear.

AI Glossary / Graphs and symbolic AI / Beginner

Satisficing

Definition

Satisficing is a decision strategy that seeks an option meeting an acceptable standard and can stop searching once such an option is found. It contrasts with requiring proof that the chosen option is the best among every available alternative.

What it means

People and AI systems rarely have unlimited time, information or computing resources. A restaurant owner might need a workable staff roster by lunchtime, even though an exhaustive search could eventually find a cheaper one. Satisficing makes the acceptance standard part of the decision procedure.

The standard is often called an aspiration level. It can be a minimum score, a maximum cost, or several requirements that must all hold. Acceptable does not have to mean mediocre: a demanding threshold can require excellent performance.

The crucial distinction is between meeting the standard and establishing optimality. An acceptable solution might happen to be optimal, but the stopping rule does not establish that fact. Conversely, stopping because time ran out does not automatically mean an acceptable solution was found.

How it works

First define the decision, the available search process and the criteria for accepting a candidate. Separate requirements that cannot be waived from preferences that allow trade-offs. A candidate that meets the price target but violates a mandatory capacity requirement should not pass.

Next generate or inspect alternatives. Evaluate each against the current aspiration level. A simple version accepts the first qualifying alternative; other procedures retain candidates or revise aspiration levels in response to experience.

The order of search can therefore affect the result. Two searches using identical requirements may stop at different acceptable options because they encounter candidates differently.

Finally record why the search stopped. Success means the acceptance criteria were met; exhaustion, timeout and unavailable information are different outcomes. Revising a threshold may be reasonable, but the revised standard and its consequences should remain visible rather than being silently substituted.

Three practical examples

A scheduling assistant accepts a roster with full coverage and overtime below a specified limit, without proving that overtime is minimal.
A document-processing team selects a configuration that meets its measured accuracy and response-time targets before a launch deadline.
A household compares replacement appliances until one meets its dimensions, price limit and required features, then ends the search.

Think of it this way

Think of looking for a parking space within a ten-minute walk of your destination. Once you find a permitted, available space inside that limit, you may park rather than inspect every nearby street for the closest possible space.

The analogy concerns a stopping rule, not the assumption that all acceptable options are equally desirable. Accessibility, permission and other essential conditions still matter, and the walking limit may need careful justification.

Worked example

Suppose an invented system configuration must achieve at least 92% accuracy on a defined evaluation set and have a median response time no greater than 300 milliseconds. The search examines candidates in this order:

A: 94% accuracy and 340 milliseconds.
B: 93% accuracy and 280 milliseconds.
C: 96% accuracy and 240 milliseconds.

A fails the time requirement. B meets both requirements, so a first-acceptable satisficing rule selects B and stops before evaluating C. If C's measurements were later known, C would look preferable on both measures, but B was still acceptable under the stated stopping rule.

A different inspection order could select C immediately. That dependence on search order is a feature of this procedure, not an arithmetic error.

These invented measurements do not establish production reliability. The evaluation set, uncertainty and whether median latency hides slow cases remain separate questions when deciding whether the thresholds are suitable.

Illustrative case study

In a fictional project, Cedar Rosters needs a weekly staffing plan before managers start their morning meetings. Its planning system searches for complete coverage with no prohibited shift combinations and overtime below a published cap.

The first feasible plan within the cap is sent for review. The report says “meets the acceptance criteria,” rather than “lowest possible staffing cost.” If no plan qualifies before the deadline, the system reports the unresolved requirements instead of quietly increasing the overtime allowance.

The case shows how satisficing can support a practical deadline while preserving an honest account of what the search achieved. It provides no evidence of actual savings.

Common misconceptions and mistakes

Satisficing means choosing randomly. Candidates are judged against an acceptance standard.
An acceptable result is necessarily optimal. The rule does not establish that comparison.
A time limit alone defines satisficing. The system also needs criteria for a satisfactory result.
Standards must remain fixed forever. They may change, but revisions need to be explicit.
Meeting a model's threshold guarantees real-world suitability. Poor measurements or missing requirements can make the threshold misleading.

Business use

Satisficing is useful when further search has a cost and an adequate decision is valuable sooner. Teams should specify who sets acceptance criteria, which requirements are mandatory and what happens when nothing qualifies.

Compare the cost of continued search with the possible benefit of improvement. An operational deadline can justify a stopping rule, but it does not justify claiming that the selected choice is globally best.

For AI-assisted decisions, retain the candidate's evaluation evidence and the stopping reason. Periodically check whether the thresholds still reflect user needs. A conveniently reachable target can otherwise become detached from the purpose of the service.

Check your understanding

A search stops after finding a plan costing 480 units, below its 500-unit acceptance limit. Has it proved that no plan costs 450 units?

Answer: No. It has established acceptability under the cost limit, not minimum cost. Additional search or an optimality argument would be needed for the stronger claim.

Related terms

Bounded rationality
Heuristic search
Optimization
Stopping rule
Search cost
aspiration level
feasible solution
decision criterion

Learn next

Choose a routine decision and write three acceptance requirements. Separate mandatory constraints from preferences. Then describe the outcomes “acceptable option found,” “search exhausted” and “time limit reached” without treating them as equivalent.

Sources

Herbert Simon's Nobel lecture, discussing search, aspiration levels and satisficing: https://www.nobelprize.org/uploads/2018/06/simon-lecture.pdf

Relevant primary-source search excerpts were reviewed; direct PDF access returned a 403 restriction. The configuration calculation and Cedar Rosters case are original teaching illustrations, not reported experiments.

Previous: Satisfiability threshold conjecture Next: Satterthwaite solution