Optimal Use of Limited Computational Resources

26 concepts17 papers1 proposal

Traditional models of rationality assume that agents possess infinite, cost-free computing power. Under this "ideal" framework, human deviations from logic—such as the anchoring bias or the St. Petersburg Paradox—are viewed as fundamental errors or "bugs" in our evolutionary programming. However, a growing body of research suggests these behaviors are not failures of logic, but rather the resource-rational output of systems optimizing for limited time, memory, and metabolic energy.

The Accuracy-Cost Trade-off

The core intuition of resource-rationality is that thinking is a physical process with an associated cost. As Lieder argues, a truly rational agent should not seek the most accurate answer possible, but rather the algorithm that achieves the best balance between solution accuracy and the computational cost of reaching it.

This can be framed as a meta-level optimization problem. Instead of simply maximizing the utility of a choice, the agent maximizes: $$Utility_{Total} = Expected\,Utility(Action) - Cost(Computation)$$

For example, the "anchoring bias"—the tendency to stay too close to an initial starting value—is often cited as a sign of human irrationality. However, Lieder et al. demonstrate that this behavior is mathematically optimal when the cost of mental adjustment is taken into account. An agent should stop adjusting their estimate the moment the marginal cost of further thinking outweighs the potential gain in accuracy. As shown by Lieder et al., what looks like "insufficient adjustment" is actually a calculated stop at the point of diminishing returns.

Biases as Efficient Algorithms

When computation is treated as a scarce resource, many "irrational" behaviors emerge as efficient algorithmic shortcuts.

  • Memory Constraints: By modeling decision-makers as finite automata with strictly limited memory, Halpern et al. show that simple, state-based strategies are near-optimal. These constraints naturally give rise to recency effects and first-impression biases, which are efficient ways to manage limited storage capacity.
  • Noisy Environments: In the context of the St. Petersburg Paradox, the authors argue that human behavior reflects an optimization for rapid estimation with minimal data. Seemingly irrational choices are actually hallmarks of algorithmic efficiency when operating under severe resource pressure.
  • Compensatory Heuristics: Sometimes one bias exists to correct for another. Guo et al. suggest that "pessimism bias" in uncertain environments may be a clever compensatory mechanism that allows simplified choice policies to rival the performance of mathematically optimal solutions.

The Emergence of Heuristics

Where do these shortcuts come from? They may be learned approximations of optimal strategies. Binz et al. used meta-learning to train neural networks under varying resource constraints, finding that the models naturally developed a spectrum of heuristics that mirror human behavior. This suggests that our "biases" are not arbitrary errors but are shaped by the inherent limitations of our biological hardware.

This perspective extends to how we perceive structure in the world. Teng et al. found that humans tend to "over-cluster" data, seeing narrow groups where only one exists. This "bias" is resource-rational because simpler structural representations are easier to track and maintain than perfectly accurate, complex ones.

Applications in AI and Mental Health

Understanding the computational logic of human behavior has significant implications for both engineering and clinical science.

In AI Alignment, Tan Zhi-Xuan et al. argue that we should move away from matching "raw" human preferences, which are often inconsistent. Instead, alignment should focus on the resource-rational heuristics humans use to manage limited information, shifting the focus from correcting "flaws" to understanding the normative standards humans negotiate under constraints.

In Clinical Psychology, the resource-rational framework offers a new way to view conditions like anxiety and depression. Rather than seeing these as simple impairments in learning, Fang et al. suggest that individuals in these states may rely more heavily on "cheap" heuristics because internal cognitive demands (like rumination) reduce the available bandwidth for complex calculations. Their decision-making is not "broken"; it is a rational response to a high-noise, low-bandwidth internal environment.

By framing the mind as a boundedly optimal system, researchers like Griffiths et al. provide a bridge between idealized mathematical models and the messy reality of biological cognition. We are not "irrational" machines; we are efficient ones, doing as well as we can with the resources we have.

Key Foundational Work

Algorithmic Rationality: Game Theory with Costly Computation (2014) — contributed 1 concept elements to this direction.
I'm Doing as Well as I Can: Modeling People as Rational Finite Automata (2012) — contributed 1 concept elements to this direction.
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 2, frozen text):

@misc{lacuna_bounded_optimality_d8da43b0_v2,
  title = {bounded optimality},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bounded-optimality/rdr_d8da43b097de45fa9194e83a1044bab5/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake