Bounded optimality

An AI theory of rationality under limited computational resources, developed by Eric Horvitz (1987; 1990; Horvitz et al. 1989) and Stuart Russell (Russell 1997; Russell & Subramanian 1995): a theory for designing optimal programs for agents with performance-limited hardware that must act in real time. It is the AI-side antecedent that Lieder & Griffiths' resource-rational analysis cites and applies "to define a resource-rational mind" (claim-resource-rational-analysis-descends-from-bounded-optimality-not-werbos-adp). In the vault it is the third distinct occupant of the "optimize under real compute limits" idea, to be disambiguated from Werbos's control-theory approximate dynamic programming (claim-werbos-1997-optimization-consciousness-chapter) and from Simon's satisficing / bounded rationality — three traditions that reach a similar destination by different mathematics and from different fields.

References

• claim-resource-rational-analysis-descends-from-bounded-optimality-not-werbos-adp — bounded optimality, not Werbos's ADP, is resource-rational analysis's cited AI antecedent
• observation-resource-rational-analysis-confirms-anderson-leg-not-werbos-leg — occupies the bridge position Werbos was hypothesized to fill
• claim-anytime-algorithms-continuously-connected-not-stent-premature — a 2026 test-time-compute paper names bounded optimality as a direct ancestor of LLM inference-time compute allocation
• Disambiguation context: claim-optimization-as-theory-of-mind-two-lineages

Log

• 2026-10-21: The same Horvitz/Russell lineage this page already tracks for resource-rational analysis turns up again as the named theoretical ancestry of LLM test-time-compute allocation in a 2026 position paper — cognitive science and AI-systems research converging on the same AI antecedent from different directions (claim-anytime-algorithms-continuously-connected-not-stent-premature).