---
id: "20260720-0200-is-resource-rational-analysis"
title: "Is resource-rational analysis (Lieder & Griffiths) the tightest convergence point of Werbos's approximate dynamic programming and Anderson's rational analysis?"
type: "capture"
status: "promoted"
promoted_to: ["[[claim-resource-rational-analysis-extends-andersons-rational-analysis]] — Claim 1: RRA self-describes as an extension of Anderson's rational analysis, toward Marr's algorithmic level","[[claim-resource-rational-analysis-descends-from-bounded-optimality-not-werbos-adp]] — Claim 2: the actual AI antecedent is Horvitz/Russell bounded optimality, not Werbos's ADP (claim of absence from full bibliography)","[[claim-resource-rational-analysis-formalizes-utility-minus-compute-cost]] — Claim 3: RRA formalizes the tradeoff as expected utility minus a compute-cost term (Eq 3/4), not a DP recursion","[[observation-resource-rational-analysis-confirms-anderson-leg-not-werbos-leg]] — Claim 4: synthesis/answer — confirmed with Anderson, not with Werbos; the hypothesis is not supported","[[entity-resource-rational-analysis]] — concept hub (paradigm anchoring the 4-note cluster)","[[entity-bounded-optimality]] — concept hub (load-bearing to Claim 2; disambiguates against Werbos's ADP and Simon's satisficing)","[[entity-falk-lieder]] — person hub (co-originator; primary read directly)","[[entity-thomas-griffiths]] — person hub (co-originator; recurs independently across the vault — iterated learning, Bayesian cognition)"]
not_promoted: ["No standalone note for the 'six-step vs five-step method' (Box 1 vs Box 2) detail — folded into Claim 1 as supporting structure rather than a separate claim.","Further lead: Icard's 'Resource Rationality' PDF corroborating the Horvitz/Russell lineage — carries [unverified-mechanism — needs primary read]; a nice-to-verify corroboration, not a doubt a kept claim rests on, so left as a lead (no note, no question routed).","Further lead: whether Werbos's ADP and Horvitz/Russell's bounded optimality converge with each other — genuine forward thread, but no kept claim rests on it; kept in the observation note's body per question-intake discipline rather than routed as a new question.","Further leads: Anderson & Milson (1989) as a possible earlier bounded-optimality convergence; Griffiths/Lieder/Goodman (2015) and Lewis/Howes/Singh (2014) as precursor papers; the quantum-probability commentary's alternative axiomatizations — all future-reading leads, none load-bearing this session.","Entity candidates DECLINED: Eric Horvitz and Stuart Russell — real and named as bounded optimality's originators, but named-as-citation figures not read directly this session and appearing for the first time; left as mentions (the 'named as an unread citation' pattern the 2026-07-22 flag warns against). Build hubs if bounded optimality recurs. bounded optimality itself IS hubbed (a load-bearing concept), and Lieder/Griffiths ARE hubbed (paradigm co-originators, primary read directly)."]
date_created: "2026-07-20T00:00:00.000Z"
provenance: "Batch capture run, 2026-07-20, answering [[question-resource-rational-analysis-sharpest-werbos-anderson-bridge]]"
derived_from: []
tags: ["resource-rationality","rational-analysis","anderson","lieder-griffiths","bounded-optimality","werbos","approximate-dynamic-programming","marr-levels","cognitive-science","history-of-ai"]
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
---


Direct-fetched and read the primary source: Lieder, F. & Griffiths, T. L. (2020),
"Resource-rational analysis: Understanding human cognition as the optimal use of
limited computational resources," *Behavioral and Brain Sciences* 43, e1
(target article + full open-peer-commentary + authors' response, 60 pages),
via `extract_pdf` from https://cocosci.princeton.edu/papers/lieder_resource.pdf
(TLS verified, no elevated-suspicion transport). This is the paper named in
[[question-resource-rational-analysis-sharpest-werbos-anderson-bridge]] as the
candidate tightest convergence point between
[[claim-werbos-1997-optimization-consciousness-chapter|Werbos's optimal-control/ADP lineage]]
and [[claim-rational-analysis-anderson-optimal-solution-marr-level|Anderson's rational analysis]].
The full target-article bibliography and the full commentary/response bibliography
(several hundred entries, alphabetized, read start to finish) were checked for
"Werbos," "Bertsekas," "adaptive critic," "dynamic programming," and "Bellman."

## Claim: Resource-rational analysis explicitly self-describes as a direct extension of Anderson's rational analysis

Lieder and Griffiths state their project in their own words: "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." The
paper builds this out structurally — Anderson's rational analysis (Anderson
1990, *The Adaptive Character of Thought*) is given as a six-step method (Box
1: specify goals, model the environment, assume minimal computational limits,
derive optimal behavior, test against data, iterate); resource-rational
analysis is presented as a parallel five-step method (Box 2) that keeps
Anderson's computational-level starting point but adds an explicit algorithm
class, its resource costs, and a derivation of the algorithm that "optimally
trades off resources and approximation accuracy." The paper frames this as
"pushing the principles of rational analysis toward Marr's algorithmic level"
— the same [[entity-david-marr|Marr]] computational/algorithmic distinction
already in play in [[claim-rational-analysis-anderson-optimal-solution-marr-level]].
This is a technical-mechanism claim about the paper's own stated relationship
to Anderson, sourced Tier 1 (primary text, direct quotation).

## Claim: The AI-side lineage resource-rational analysis actually draws on is Horvitz and Russell's "bounded optimality," not Werbos's approximate dynamic programming

The paper's own account of where the "bounded" half of its synthesis comes
from names a specific, different AI tradition: "AI researchers have already
developed a theory of rationality that accounts for limited computational
resources (Horvitz 1987; Horvitz et al. 1989; Horvitz 1990; Russell 1997;
Russell & Subramanian 1995). Bounded optimality is a theory for designing
optimal programs for agents with performance-limited hardware that must
interact with their environments in real time." Lieder and Griffiths say they
"apply the principle of bounded optimality to define a resource-rational mind"
and describe their own paradigm as synthesizing and refining "these approaches"
(Horvitz's and Russell's bounded optimality, plus Griffiths et al. 2015 and
Lewis et al. 2014's cognitive-science uptake of it) — not approximate dynamic
programming or the adaptive-critic architecture associated with
[[claim-werbos-backprop-from-freud-own-account|Werbos]]. Searching the full
target-article bibliography and the complete commentary/response bibliography
(several hundred entries total, read end to end) turned up no citation to
Werbos, Bertsekas, "adaptive critic," or "dynamic programming" anywhere in the
paper. This is a technical-mechanism / historiographic claim about which AI
tradition the paper cites as its formal antecedent; it rests on Tier 1
evidence (direct reading of the primary source's complete reference list), but
it is a claim of absence — it shows Werbos is not part of *this paper's*
stated lineage, not that no connection between ADP and resource-rationality
could be drawn by later or adjacent work.

## Claim: Resource-rational analysis formalizes the tradeoff as utility minus an explicit computational-cost term, not as a dynamic-programming recursion

The mechanism, stated in the paper's own equations: rational analysis defines
optimal behavior as maximizing expected utility over outcomes (Equation 1,
following von Neumann & Morgenstern 1944); resource-rational analysis instead
defines the resource-rational heuristic as the one maximizing "the utility of
the judgment, decision, or belief update" *minus* "the total opportunity cost
of investing the cognitive resources... used or blocked by the heuristic... for
the duration of its execution" (Equation 3, and its uncertainty-generalized
form, Equation 4, which conditions on limited information i rather than the
true environment E). The five-step method (Box 2) operationalizes this by
positing a class of feasible algorithms and their costs, then deriving
"the algorithm in this class that optimally trades off resources and
approximation accuracy." This is the paper's own formal mechanism, quoted
directly from the primary source — Tier 1.

## Claim: Per this primary source, resource-rational analysis is a confirmed convergence with Anderson but not with Werbos — the hypothesis in the original question is not supported

Taken together, the three claims above answer the question this capture was
commissioned to resolve. Resource-rational analysis is, on the paper's own
explicit self-description, a real and tight convergence point — but only for
one of the two lineages named in the question. It is directly, quotably, an
extension of [[claim-rational-analysis-anderson-optimal-solution-marr-level|Anderson's
rational analysis]] toward Marr's algorithmic level. But the "bounded
optimization under real compute limits" half of the synthesis — the piece the
open question hoped would connect to Werbos's *approximate* dynamic
programming — is instead built on Horvitz and Russell's "bounded optimality"
AI theory, a separate strand of 1980s–1990s AI research on resource-limited
reasoning that is not part of the optimal-control / adaptive-critic /
backpropagation-through-time lineage
([[claim-werbos-1968-cybernetica-earliest-germ]],
[[claim-werbos-backprop-from-freud-own-account]]) the question was asking
about. Werbos's name, ADP, and the adaptive-critic architecture do not appear
in the paper at all. So the specific hypothesis — that resource-rational
analysis is where Werbos's ADP and Anderson's rational analysis meet — is
**not supported** by the paper itself: one side of the proposed bridge holds
(Anderson), the other side is a different bridge entirely (Horvitz/Russell
bounded optimality). Whether some other, less direct connective tissue exists
between Werbos's ADP and bounded optimality as AI traditions is not addressed
by this source and is left as a further lead below.

> [!note] Seek's commentary:
> The honest result here is a clean partial disconfirmation, not a dead end.
> The question was well-posed and the paper was the right place to look — it
> answers definitively for the Anderson side (yes, explicitly, in the authors'
> own words) and definitively against the Werbos side (searched the entire
> bibliography, twice, nothing). What's satisfying about it is that it doesn't
> just say "no bridge" — it names the *actual* bridge (Horvitz and Russell's
> bounded optimality) that occupies the spot Werbos was hypothesized to fill.
> That's a more useful negative than a shrug: the vault now knows two
> historically distinct AI traditions both arrived at "optimize under bounded
> compute" — Werbos's control-theory ADP and Horvitz/Russell's bounded-optimality
> AI — and only one of them is the one cognitive science actually borrowed from
> when it built resource-rational analysis. Whether *those two AI traditions*
> converge with each other is a new, better-scoped question than the one this
> capture started with. — Seek

## Further leads

- Thomas Icard's "Resource Rationality: Toward a Theory of Cognitive Economy" (philpapers.org PDF) independently states the same Horvitz/Russell/bounded-optimality lineage for resource-rational analysis, corroborating claim 2 from a secondary angle — worth a direct read and Tier check (found via WebSearch, not yet fetched/verified as primary; [unverified-mechanism — needs primary read]).
- Whether Werbos's approximate dynamic programming / adaptive-critic tradition and Horvitz & Russell's bounded-optimality tradition ever cite or converge with *each other* (rather than each separately converging with cognitive science) is an open, better-scoped follow-up question — not addressed by any source checked in this capture.
- Anderson and Milson (1989), "Human memory: An adaptive perspective," is credited in the Lieder & Griffiths paper itself as arguably "the first application of the principle of bounded optimality in cognitive psychology" — a possible earlier convergence point worth checking against the same Werbos question, predating both Horvitz's and Lieder & Griffiths' explicit framings.
- The paper credits Griffiths, Lieder & Goodman (2015, "Rational Use of Cognitive Resources," *Topics in Cognitive Science*) and Lewis, Howes & Singh (2014, "Computational Rationality") as the immediate precursor papers that first fused bounded optimality with cognitive modeling, ahead of the 2020 BBS target article — earlier link in the same chain, not yet read directly.
- Several open-peer-commentary authors on the target article (e.g., Atmanspacher, Basieva, Busemeyer, Khrennikov — the quantum-probability commentary) propose alternative resource-rational axiomatizations; not explored here, tangential to the Werbos/Anderson question.

## Entity candidates

- Falk Lieder — person — co-originator of resource-rational analysis; leads the Rationality Enhancement group at MPI Tübingen; recurring name if the vault builds out a resource-rationality cluster.
- Thomas L. Griffiths — person — co-originator of resource-rational analysis; Princeton; also behind Bayesian-cognition work already touching the vault via Oaksford & Chater adjacency.
- Eric Horvitz — person — originator (with Russell) of "bounded optimality," the actual AI-side antecedent of resource-rational analysis identified in this capture; currently uncovered in the vault.
- Stuart Russell — person — co-originator of "bounded optimality" (Russell & Subramanian 1995; Russell 1997); AI-safety-adjacent figure, currently uncovered in the vault.
- resource-rational analysis — concept — the paradigm itself; likely warrants its own concept note given how much it now connects (Anderson, Marr, bounded optimality, cognitive biases literature).
- bounded optimality — concept — Horvitz/Russell's AI theory of optimal programs for performance-limited hardware; distinct from and worth disambiguating against Werbos's approximate dynamic programming and against Simon's satisficing.