Does limited rationality mean irrationality?

No. It concerns making appropriate decisions under practical limits rather than assuming unlimited computation.

Is more reasoning always beneficial?

No. Its expected benefit can be smaller than its computational or delay cost.

Is it identical to bounded optimality?

No. Bounded optimality is a more specific formal approach to evaluating agent programs under architectural constraints; limited rationality is a broader idea.

AI Glossary / Graphs and symbolic AI / Beginner

Limited rationality

Definition

Limited rationality is decision-making that accounts for limits on time, information, memory, or computation. In AI, it recognizes that an agent often cannot calculate the theoretically best action by exhaustively examining every possibility. The design challenge is to act appropriately with available resources, including deciding how much effort to spend on further reasoning.

What it means

An idealized decision rule may assume that all relevant calculations can be completed before acting. A real system must operate on finite hardware and often faces a deadline. Spending too long searching for a better answer can make the eventual answer less useful or too late.
Limited rationality does not mean arbitrary behavior or a rejection of careful reasoning. It changes what must be included in the assessment: the cost and timing of the reasoning process matter alongside the quality of the selected action. The concept overlaps with bounded rationality, while particular theories use more precise definitions that should not be treated as interchangeable without explanation.

How it works

An agent can use heuristics, restricted search, approximate inference, or algorithms that improve their current answer as more time becomes available. It may also compare the expected benefit of another computation with its cost, including delay.
This introduces a decision about computation itself. The agent needs some way to estimate whether additional analysis is likely to change the action enough to matter. Such estimates can be imperfect and costly, so the approach does not magically remove all resource limits. A practical design defines a resource budget, an acceptable action policy, and a stopping or fallback rule appropriate to the task.

Three practical examples

A game program searches promising continuations within its move deadline rather than enumerating the entire game tree.
A scheduling tool returns a feasible plan quickly and improves it while more computation remains available.
A response system chooses whether another information-gathering step is likely to improve an answer enough to justify additional delay.

Think of it this way

Think of choosing a route while a departure time approaches. Comparing more alternatives may help, but continued deliberation can itself consume the time available for travel. The analogy is limited because formal AI models must define costs, uncertainties, and possible computations explicitly; ordinary intuition does not provide a universally correct stopping rule.

Worked example

Consider an invented decision process using abstract utility units. Acting now with the best available option has expected utility 100. An additional computation is estimated to have a 0.25 probability of revealing an option worth 120; otherwise, the current option remains available with value 100. Assume these are the only possible outcomes and that the estimates are correct for this illustration.
Before computation costs, the expected value after the extra analysis is 0.25 times 120 plus 0.75 times 100 = 105. The expected improvement is therefore 5 utility units.
If the computation and associated delay cost 3 utility units, its net expected value is 105 - 3 = 102, exceeding the value 100 from acting now. Under this simplified one-step comparison, additional analysis is worthwhile. If the cost is instead 7, the net value is 98, so acting now is preferred.
This calculation does not say that faster is always better or that the agent knows the value of computation perfectly. It also ignores the possibility of a sequence of further computations. It demonstrates why an analysis step can be useful at one cost and undesirable at another, even when its potential improvement is unchanged.

Illustrative case study

In a fictional planning project, Hawthorn Schedule spends most of its allowed time seeking a small improvement over a feasible plan. Sometimes it reaches the deadline without returning the plan already available. The team introduces a mechanism that preserves a usable incumbent and measures quality at several time limits.
It evaluates whether further search tends to improve the operational result enough to justify the delay. The invented case illustrates resource-aware design; it does not establish a universal timeout or claim that every approximate plan is acceptable.

Common misconceptions and mistakes

Limited rationality is not the same as knowingly ignoring essential constraints. A faster answer can still be invalid or unusable. The quality requirements of the task remain part of the decision problem.
Do not equate it with satisficing in every context. Satisficing is one possible approach, while limited-resource decision-making includes many strategies. Also, an algorithm described as approximate is not automatically well adapted to a deadline. Its errors, runtime behavior, and fallback conditions need evaluation. Finally, better hardware can change the feasible computations without eliminating all tradeoffs between analysis and action.

Business use

The concept helps teams define what “good performance” means for an AI system operating under real constraints. Include latency, available information, and computation cost alongside answer quality. Establish which requirements are mandatory and which aspects can improve progressively.
Evaluate the system under realistic workloads and deadlines, including difficult cases. Measure the quality of the answer available when the decision must actually be made. When using stopping rules, document their assumptions and failure behavior. This makes resource limits an explicit design consideration rather than an excuse introduced after the system fails to deliver a useful result.

Check your understanding

Why is the extra computation rejected when its cost is 7 in the worked example?

Answer: Its expected improvement is only 5. Subtracting the cost from the expected post-analysis value gives 98, below the value 100 of acting immediately under the stated assumptions.

Related terms

Learn next

List the costs of delaying a decision in a familiar workflow. Compare a quick feasible answer with a slower improved answer using an explicitly stated objective. Identify which assumptions would need evidence before a value-of-computation calculation could guide the real system.

Sources

Sources: Russell and Norvig, Artificial Intelligence: A Modern Approach, original introduction, limited rationality: https://aima.cs.berkeley.edu/~russell/intro.html

Previous: Limited Memory AI Next: Linda-like systems