Resource-Rational Modeling of Human Decision Making

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 222 papers28 conceptsfrozen 2026-10-04

For decades, behavioral economics and cognitive science have cataloged a long list of human "irrationalities"—systemic deviations from the predictions of classical logic and expected utility theory. We succumb to the anchoring bias, we are over-optimistic in some scenarios and pessimistic in others, and we fail to solve even simple paradoxes like the St. Petersburg Paradox.

However, a growing body of research suggests that these behaviors are not "bugs" in our biological software. Instead, they are the optimal outputs of a system designed to solve a specific problem: maximizing utility under strict computational constraints. This perspective, often termed resource-rationality, shifts the focus from what we decide to how we must compute those decisions given limited time, memory, and energy.

The Utility of Thinking

Traditional rational models assume that agents have infinite, free computing power. In Algorithmic Rationality: Game Theory with Costly Computation, Joseph Halpern and Rafael Pass argue that this assumption is fundamentally flawed. In the real world, computation is a costly resource.

A resource-rational agent does not simply maximize the expected reward of an action; they maximize the expected reward minus the cost of the computation required to choose that action. As Thomas L. Griffiths notes in Understanding Human Intelligence through Human Limitations, human intelligence is defined by its constraints—limited time, limited computation, and limited communication. Our cognitive heuristics are actually "rational metareasoning" strategies that optimize this trade-off.

Biases as Optimal Shortcuts

When we view the mind as an optimizer of limited resources, classic cognitive biases begin to look like efficient algorithmic responses.

Anchoring and Adjustment

The anchoring bias—our tendency to stay too close to an initial starting value—is often cited as a failure of reasoning. However, Falk Lieder, Quentin J. M. Huys, and Noah D. Goodman demonstrate in The anchoring bias reflects rational use of cognitive resources that this is a resource-rational strategy.

If every second of "thinking" has a metabolic or temporal cost, it is mathematically optimal to stop adjusting your estimate once the marginal cost of further deliberation exceeds the expected gain in accuracy. We aren't "lazy"; we are calculating the point of diminishing returns.

Pessimism and Exploration

Similarly, the "pessimism bias" in uncertain environments can be seen as a compensatory mechanism. Dalin Guo and Angela J. Yu show in Why so gloomy? that combining a simplified choice policy with a pessimistic volatility assumption allows humans to rival the performance of mathematically optimal solutions while using far less processing power.

Modeling the Bounded Agent

To capture these dynamics, researchers have moved toward modeling humans as finite automata—machines with a strictly limited number of internal states.

In I'm Doing as Well as I Can: Modeling People as Rational Finite Automata, Joseph Y. Halpern and Lior Seeman show that state-based strategies are near-optimal for memory-constrained agents. These models naturally give rise to recency and first-impression effects. Further work by Joseph Halpern and Xinming Liu in Bounded Rationality in Las Vegas demonstrates that even complex behaviors like optimism and negativity bias emerge as rational adaptations when agents are modeled as Probabilistic Finite Automata (PFAs) playing multi-armed bandit problems.

This shift is also visible in how we think about task decomposition. Carlos G. Correa and Mark K. Ho argue in Humans decompose tasks by trading off utility and computational cost that we don't just choose subgoals to make a path shorter; we choose them to make the process of thinking easier.

Implications for AI and Alignment

Understanding resource-rationality is not just a matter of psychological curiosity; it is a prerequisite for building AI that can interact safely with humans.

If an AI assumes a human is a perfect utility maximizer, it will systematically misinterpret human actions. As Sunayana Rane and Mark K. Ho explain in Concept Alignment as a Prerequisite for Value Alignment, humans use "construals"—simplified mental representations—to manage cognitive load. An AI that fails to account for these shortcuts will infer the wrong goals from human behavior.

By framing human "errors" as resource-rational optimizations, we move away from trying to "fix" human bias and toward a framework where AI can align with the actual, computationally-constrained ways that humans represent and solve problems in the real world.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 17 papers
  2. Version 1 · 2026-10-04 · 22 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_bounded_optimality_d8da43b0,
  title = {bounded optimality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bounded-optimality/rdr_d8da43b097de45fa9194e83a1044bab5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_resource_rational_modeling_of_human_deci_d8da43b0_v1,
  title = {Resource-Rational Modeling of Human Decision Making},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/resource-rational-modeling-of-human-decision-making/rdr_d8da43b097de45fa9194e83a1044bab5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}