Is bounded rationality only about human psychology?

No. It also informs the design of artificial agents that reason under computational and information limits.

Does accepting an adequate option prove it is optimal?

No. Adequacy depends on the acceptance rule; optimality requires a stronger comparison.

Can a more complex model make decisions worse?

It can if added computation or information demands interfere with timely, dependable action, though this must be assessed in context.

AI Glossary / Graphs and symbolic AI / Intermediate

Bounded rationality

Definition

Bounded rationality is decision making shaped by limits on information, time, attention, memory, or computation. A person or AI agent cannot always examine every option and calculate every consequence. Instead, the decision process must work within available resources. The concept explains why practical reasoning may use simplified models, selective search, or acceptable stopping rules.

What it means

An idealized decision maker might compare all possible actions and choose the one with the greatest expected benefit. Real decision makers must also decide what to investigate, how much effort to spend, and when to act. Those choices can affect the outcome as much as the final comparison of alternatives.
Bounded rationality is a broad perspective, not a single algorithm or an excuse for careless decisions. A procedure can respect resource limits while still being evaluated for accuracy, fairness, reliability, and usefulness. Different theories describe these limits in different ways, so a claim that a system is bounded-rational should explain what is bounded and how.

How it works

A practical procedure starts with a decision objective and the information already available. It searches a limited set of alternatives, uses a representation simple enough to handle, and applies a stopping rule. One familiar approach is satisficing: accepting an option that meets a stated aspiration level instead of establishing that no better option exists.
Other approaches allocate computation according to its expected benefit or explicitly constrain the information a decision process can use. These are different models of resource-limited reasoning. Their common concern is that deliberation itself has costs and constraints. Choosing a useful stopping rule still requires assumptions about the task and the consequences of an inadequate answer.

Three practical examples

A scheduling assistant examines promising timetable changes until its response deadline, then returns the best feasible arrangement found.
A buyer uses a shortlist of suppliers that meet essential requirements because investigating every supplier would consume disproportionate effort.
An embedded robot uses a compact local map and a limited planning horizon because its processor cannot repeatedly solve a complete world model.

Think of it this way

Think of selecting a meal before a short lunch break ends. You might choose the first suitable option rather than compare every restaurant in the city. The analogy illustrates limited search and an acceptance threshold. It does not imply that every quick choice is wise, or that a casual lunch rule is appropriate for a high-consequence engineering decision.

Worked example

Suppose a fictional planning assistant has 30 seconds and needs 10 seconds to evaluate each candidate. A predetermined sequence contains plans with scores 62, 78, 85, and 93; higher is better. Its acceptance threshold is 75.
A satisficing rule stops at the second plan, after 20 seconds, with score 78. A best-within-budget rule evaluates three plans and returns 85 after 30 seconds. Exhaustive comparison would select 93 but requires 40 seconds and misses the deadline. These are different procedures under the same resource limit. The example does not prove that either practical rule is universally best. If acting 10 seconds earlier has value, or the third evaluation sometimes takes longer, the preferred procedure could change. A score threshold also needs justification: meeting 75 does not establish compliance with requirements that the score fails to represent.

Illustrative case study

At fictional Willow Events, an assistant proposes room assignments for a conference. The first design searches extensively for small improvements and sometimes delivers its answer after staff need to publish the schedule. The team introduces an explicit planning budget and a rule for returning a valid provisional schedule.
Staff compare the resulting schedules and response times across representative scenarios, including difficult cases. They retain a process for escalating unresolved conflicts. The fictional case illustrates adapting reasoning to operating constraints; it does not claim a measured improvement or suggest that time pressure makes invalid assignments acceptable.

Common misconceptions and mistakes

Bounded rationality does not mean irrationality. It recognizes that reasoning has limits, although a particular decision can still be mistaken or poorly justified. Nor is it synonymous with satisficing: satisficing is one possible response to limited resources.
It also differs from bounded optimality, which makes a stronger claim about the best feasible agent program for a specified architecture and environment. Bounded suboptimal search is different again, typically providing a solution-cost guarantee. Finally, adding computing power does not automatically resolve missing information, unclear objectives, or a misleading model of the decision.

Business use

For businesses, the concept makes decision procedures easier to assess realistically. Teams can state which alternatives are considered, what information is omitted, how long analysis may take, and who handles unresolved cases. This helps distinguish deliberate simplification from accidental neglect. It also supports evaluating an AI assistant by the usefulness of timely decisions, while preserving essential constraints and documenting what its limited search cannot establish.

Check your understanding

An assistant checks five acceptable plans and stops because its time budget expires. Can it claim to have found the best possible plan?

Answer: Only if additional reasoning establishes that claim. The time-limited search normally supports a narrower statement about the plans it examined.

Related terms

Learn next

Compare an aspiration-based stopping rule with a fixed-time search rule on a small decision problem. Record answer quality and decision time separately. Then identify any essential conditions that must hold regardless of the score. This shows how resource limits influence procedure without removing responsibility for the result.

Stopping rule

Sources

Sources: Ortega and colleagues, Information-Theoretic Bounded Rationality, original research: https://arxiv.org/abs/1512.06789 and https://arxiv.org/pdf/1512.06789 . Russell and Subramanian, Provably Bounded-Optimal Agents: https://www.cs.cmu.edu/afs/cs/project/jair/pub/volume2/russell95a.pdf . Simon, Nobel lecture, historical source listed for further reading: https://www.nobelprize.org/prizes/economic-sciences/1978/simon/lecture/ .

Previous: Bounded optimality Next: Bounded suboptimal search