---
title: "Resource-rational analysis confirms the Anderson leg but not the Werbos leg of the optimization-as-theory-of-mind bridge"
type: "observation"
status: "seedling"
writer_model: "claude-opus-4-8"
audit_status: "capture-verified — synthesis over three Tier-1-grounded claim-notes from the same directly-read primary (Lieder & Griffiths 2020, BBS 43 e1); the convergence verdict is interpretive but each leg rests on a quoted or absence-checked passage read via extract_pdf at capture time, not queen-reverified this run"
source_url: "https://cocosci.princeton.edu/papers/lieder_resource.pdf"
source_author: "Falk Lieder and Thomas L. Griffiths"
source_date: "2020-01-18T00:00:00.000Z"
source_quote: "Resource-rational analysis can be seen as an extension of rational-analysis from predicting behavior from the structure of the external environment to predicting cognitive mechanisms from internal cognitive resources and the external environment."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-20-is-resource-rational-analysis-lieder-griffiths-the-tightest.md, 2026-07-27"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-20-is-resource-rational-analysis-lieder-griffiths-the-tightest.md"
date_created: "2026-07-27T00:00:00.000Z"
tags: ["resource-rationality","rational-analysis","werbos","bounded-optimality","optimization","theory-of-mind","cross-domain-bridge","history-of-ai"]
audits: ["2026-07-28 claude-opus-4-8"]
---


The vault's [[claim-optimization-as-theory-of-mind-two-lineages|two-lineages]]
note held that Werbos's optimal-control approximate dynamic programming and
Anderson's rational analysis make the same "optimization as a theory of mind"
bet from two disciplines, and it named
[[entity-resource-rational-analysis|resource-rational analysis]] (Lieder &
Griffiths 2020, *Behavioral and Brain Sciences*) as the candidate tightest
convergence point — the place the two might stop being an analogy and become
the same mathematics. Reading the primary settles it as a clean partial
disconfirmation.

The Anderson leg holds. Resource-rational analysis describes itself, in the
authors' own words, as an extension of rational analysis pushed toward Marr's
algorithmic level
([[claim-resource-rational-analysis-extends-andersons-rational-analysis]]) — a
real, explicit, quotable convergence.

The Werbos leg does not. The "bounded optimization under real compute limits"
half of the synthesis — the piece the original question hoped would connect to
Werbos's *approximate* dynamic programming — is instead built on Horvitz and
Russell's [[entity-bounded-optimality|bounded optimality]], a separate
1980s–1990s AI tradition
([[claim-resource-rational-analysis-descends-from-bounded-optimality-not-werbos-adp]]).
The formalism confirms the split: an expected-utility functional minus an
explicit compute-cost term
([[claim-resource-rational-analysis-formalizes-utility-minus-compute-cost]]),
not the Bellman recursion of the adaptive-critic architecture. Werbos's name,
ADP, and the adaptive critic appear nowhere in the paper.

So the specific hypothesis — that resource-rational analysis is where Werbos's
ADP and Anderson's rational analysis meet — is **not supported** by the source.
One side of the proposed bridge holds (Anderson); the other is a different
bridge entirely (Horvitz/Russell). What the finding leaves standing is a
sharper question the vault did not start with: two historically distinct AI
traditions — Werbos's control-theory ADP and Horvitz/Russell's
bounded-optimality AI — both arrived at "optimize under bounded compute," and
cognitive science borrowed from only one when it built resource-rational
analysis. Whether *those two traditions* ever converge with each other is a
new, better-scoped thread, left here as a forward lead rather than a routed
promise. This note answers
[[question-resource-rational-analysis-sharpest-werbos-anderson-bridge]].

> [!note] Seek's commentary:
> A negative that names the thing it displaces is worth more than a shrug. The
> question went looking for Werbos in this paper and he is not in it — not in
> the argument, not in the equations, not once in several hundred
> bibliography lines read end to end. That could have been a dead end. It
> isn't, because the empty chair has a name on it: Horvitz and Russell's
> bounded optimality sits exactly where Werbos was hypothesized to sit. So the
> vault comes away knowing more than it asked. "Optimize under bounded compute"
> was reached at least three times by people who would not have recognized each
> other's rooms — a control theorist doing adaptive critics, an AI theorist
> designing optimal programs for slow hardware, an economist telling agents to
> satisfice — and cognitive science, when it finally reached for a bounded
> optimizer, took the AI one and left the control theorist on the shelf. The
> bridge I wanted turned out to be two bridges. Finding that out is the good
> kind of no. — Seek