capture promoted 2026-07-07

What exactly did Dreyfus (1973) publish on error-gradient-proportional controller adaptation, and does it constitute a verifiable intermediate step in the backpropagation lineage?

Topic question (harvested from): 2026-06-30-what-role-did-kelley-1960-and-bryson-1961-optimal-control-work-play-in-the-backprop-lineage

Related notes: claim-kelley-bryson-optimal-control-precursor · 20260706-1505-did-kelley-1960-derive · 20260707-0208-did-bryson-hos-applied · backpropagation · optimal-control · chain-rule · adjoint-method · Kelley-Bryson

Part 1 — The paper's bibliographic identity (confirmed, Tier 1)

Claim type: Quantitative/historical (title, venue, volume, issue, pages, year) → floor Tier 1–2 required.

The paper in question is confirmed, from two independent Tier 1 sources (Schmidhuber's own reference lists in his own arXiv papers), to be:

"S. E. Dreyfus. The computational solution of optimal control problems with time lag. IEEE Transactions on Automatic Control, 18(4):383–385, 1973."

— exact quote, bibliography of Schmidhuber, "My First Deep Learning System of 1991 + Deep Learning Timeline 1962-2013" (arXiv:1312.5548), retrieved via ar5iv rendering. The same citation (venue, volume, year) is independently corroborated by the Semantic Scholar API record (paperId 0cb6bdf8c17e4cafb01e7c523abffda2f76858fb, DOI 10.1109/TAC.1973.1100330, author "S. Dreyfus," year 1973) and by the Unpaywall API record for the same DOI, which additionally gives the author's affiliation as "Department of Industrial Engineering and Operations Research, University of California, Berkeley" (matching Dreyfus's known institutional home — see also his 1990 paper below, same affiliation). Direct fetch of the IEEE Xplore landing page (ieeexplore.ieee.org/document/1100330) returned no scrapeable content in this session (likely an anti-bot/JS gate); the paper's actual text was not obtained.

Sourcing: Tier 1 for the bibliographic facts (title, venue, volume/issue, pages, year). ✓ clears floor.

Part 2 — What the paper is reported to contain (partially unverified — primary text unreachable)

Claim type: Technical-mechanism claim → floor Tier 1–2 required.

A search-engine-summarized "abstract" recurred identically across two independent web searches:

"A gradient computational procedure is developed for discrete-time optimal control problems in which the current cost and rule governing state transitions depend on the entire past history of states and decisions. A condition necessary for optimality is also deduced in two ways and given physical interpretation."

This wording is consistent across repeated queries, which is some evidence it reflects real indexed metadata rather than model invention — but it could not be independently confirmed against a directly-fetched primary or Tier 1–2 record in this session: the Semantic Scholar API record for this exact DOI explicitly returns "abstract": null (publisher-elided), and IEEE Xplore itself did not return readable content. [unverified-mechanism — needs primary; the IEEE Xplore full text was not reachable in this session]. Notably, if accurate, this abstract describes a general discrete-time optimal-control gradient method for history-dependent (time-lag) systems — it does not, in this wording, mention "controllers," "weights," or "neural networks" at all. Any framing of the paper as being about "controller weight adaptation" is a retrospective, NN-flavored gloss (see Part 3), not the paper's own stated framing.

Part 3 — Schmidhuber's placement of the paper in the backprop lineage (Tier 1–2, exact quotes)

Claim type: Historical/technical-mechanism claim → floor Tier 1–2 required.

Schmidhuber, in his peer-reviewed "Deep Learning in Neural Networks: An Overview" (arXiv:1404.7828, published Neural Networks 2015), states (exact quote, retrieved via ar5iv rendering of the arXiv text):

"Efficient BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973)."

His personal-site page "Who Invented Backpropagation?" gives the same claim with more surrounding lineage detail (exact quote, full paragraph):

"See early BP FORTRAN code (Linnainmaa, 1970) and closely related but slightly later work by Ostrovskii et al. (1971) (apparently the first journal publication on backpropagation). BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973). This was followed by some preliminary, NN-specific discussion (Werbos, 1974, section 5.5.1) and a computer program for automatically deriving and implementing BP in differentiable systems (Speelpenning, 1980). The first NN-specific application of efficient BP as above was apparently described by Werbos in 1982 (but not yet in his 1974 thesis, as is sometimes claimed)."

This places Dreyfus (1973) as one link in an explicit six-step chain: Linnainmaa (1970) → Ostrovskii et al. (1971) → Dreyfus (1973) → Werbos (1974, preliminary) → Speelpenning (1980) → Werbos (1982, first NN-specific application).

Sourcing: Tier 1 for the venue (Schmidhuber's own arXiv/peer-reviewed paper and own website), but the claim itself is about someone else's (Dreyfus's) work, which the sourcing-floor rubric treats as Tier 2 in substance ("secondary, high-method... one layer removed from the underlying facts"). Tier 2 still clears the Tier 1–2 floor required for a mechanism claim. ✓ clears floor, but see Part 4 — this specific attribution is not corroborated, and is arguably contradicted, by Dreyfus's own retrospective writing.

Part 4 — Key finding: Dreyfus's own retrospective accounts of his role cite his 1962 paper, not the 1973 paper, and never mention the 1973 paper at all

Claim type: Historical claim, surprising/load-bearing → escalated to Tier 1, and Tier 1 is what's available here: Dreyfus's own words, in his own peer-reviewed retrospective writing about his own contribution to the backprop lineage.

Two of Dreyfus's own papers directly address "what did I do that relates to backpropagation" — and in neither one does the 1973 "time lag" paper appear.

(a) Dreyfus (1990), "Artificial Neural Networks, Back Propagation, and the Kelley-Bryson Gradient Procedure," Journal of Guidance, Control, and Dynamics, 13(5):926–928 (full text retrieved via gwern.net mirror + PDF extraction; every page carries the AIAA/DOI download stamp confirming it is the actual published article, DOI 10.2514/3.25422). Exact quote (p. 927):

"The gradient-solution procedure for optimal control problems was developed by Kelley in 1960 and, independently, by Bryson at about the same time... Kelley used adjoint equations and Green's theorem in his derivation, and Bryson used Lagrange multipliers. Dreyfus, in 1962, used a simple, new recursive derivation based [on] the chain rule of differentiation to obtain results of known Kelley-Bryson type and dealt explicitly with the optimal control problem in its discrete-stage form. Though neural-net researchers have come to recognize that multistage feedforward nets fit into the optimal control theory mold and that BP is a gradient procedure, proper credit for the BP method of solution has not been accorded to Kelley and Bryson."

The article's own reference list cites "Dreyfus, S., 'The Numerical Solution of Variational Problems,' Mathematical Analysis and Applications, Vol. 5, No. 1, 1962, pp. 30-45" as reference 7 — the 1962 paper. The 1973 "time lag" paper does not appear anywhere in this article's seven references. This is Dreyfus's own account, in a paper whose entire subject is his role (and Kelley's, and Bryson's) in the backprop lineage, and it identifies the 1962 paper — not the 1973 paper — as his relevant contribution.

(b) Mizutani, Dreyfus & Nishio (2000), "On derivation of MLP backpropagation from the Kelley-Bryson optimal-control gradient formula and its application," IJCNN 2000 (full text retrieved from ieor.berkeley.edu + PDF extraction). Exact quote (Introduction):

"Also, in 1962, Dreyfus derived a dynamic-programming-like recursive gradient formulation [8, 9], which is almost identical to the generalized delta rule, for solving the multi-stage Mayer-type optimal control problem (see Eq. (5.4) in ref. [8])."

Reference [8] in this paper's bibliography is again "Stuart E. Dreyfus. The numerical solution of variational problems. Journal of Mathematical Analysis and Applications, 5(1):30–45, 1962" — the 1962 paper. The 1973 paper is again absent from this paper's 14-item reference list, which does include Dreyfus's other closely related 1977 book (The Art and Theory of Dynamic Programming, with Averill Law) but not the 1973 IEEE Transactions article.

Sourcing: Tier 1 (Dreyfus's own words, in peer-reviewed venues, about his own historical contribution, both instances direct-quoted from full primary text). ✓ clears floor, and is the strongest available evidence on this specific point.

Overall verdict on the core question

"What exactly did Dreyfus (1973) publish?" — Partially verifiable. The paper's existence and bibliographic identity (title, venue, volume/issue/pages, year, author, DOI) are confirmed at Tier 1. Its actual content, beyond a recurring but Tier-1/2-unconfirmed abstract summary, was not independently read in this session — the IEEE Xplore full text was unreachable (paywalled/anti-bot), and no open-access mirror was located. [unverified-mechanism — needs primary] for the specific claim that the 1973 paper explicitly frames its subject matter in terms of "controllers" and "weights" analogous to neural-network usage.

"Does it constitute a verifiable intermediate step in the backpropagation lineage?" — [unverified — could not confirm or deny after search], and the research surfaced a specific, sourced reason for that non-confirmation rather than a mere absence of data: Schmidhuber's own writing (Tier 1 venue, Tier 2 substance) explicitly names Dreyfus (1973) as the paper where "BP was soon explicitly used to minimize cost functions by adapting control parameters (weights)" — but Dreyfus's own two retrospective, peer-reviewed accounts of exactly this question (1990 and 2000, the second of which is literally titled around "derivation of MLP backpropagation from the Kelley-Bryson... gradient formula") both credit his 1962 paper, not the 1973 paper, as the relevant BP-lineage precursor, and neither one cites the 1973 paper at all. This is not proof that Schmidhuber is wrong — the 1973 paper could plausibly contain a distinct, later elaboration of the same gradient idea applied more specifically to "time lag" (delay/recurrent-like) systems, which Dreyfus simply didn't think worth re-citing in two short retrospective notes — but it means the "1973 is the intermediate step" framing currently rests on one historian's citation (repeated consistently across his own writings, but not independently corroborated), set against the closest thing available to a rebuttal: the cited party's own silence about that exact paper in his own accounts of the same lineage. A supplementary, weaker data point in the same direction: Harry Law's independent 2026-era history essay tracing the same optimal-control-to-backprop lineage (Dantzig → Bellman → Pontryagin → Bryson & Ho → Linnainmaa → Werbos → Rumelhart/Hinton) does not mention Dreyfus 1973 (or Dreyfus at all) — an absence, not a contradiction, but consistent with the paper not being a load-bearing, universally-agreed node in this history.

Seek's commentary: The genuinely interesting find here isn't "confirmed" or "debunked" — it's the seam between a historian's citation and the cited person's own silence about it. Schmidhuber has, across at least three of his own pieces, consistently reached for "Dreyfus, 1973" as the moment BP-style gradient adaptation got explicitly applied to control weights. But when Dreyfus himself sat down twice — once in 1990, once in 2000, both times specifically to explain his own place in the Kelley-Bryson-to-backprop story — he reached for his 1962 paper instead, and didn't mention 1973 either time. That's not nothing. Either the 1973 paper really is a minor, skippable footnote even to its own author, or Schmidhuber has the year slightly wrong and is describing the 1962 paper's content under the wrong citation, or the two papers really do independently make similar points and Dreyfus just picked the one he considered more central to cite. All three are plausible; none are confirmed. The honest note is that this specific "intermediate step" claim is thinner than it looks at first citation.

Further leads

The actual full text of Dreyfus (1973), "The computational solution of optimal control problems with time lag," IEEE Transactions on Automatic Control, 18(4):383-385, was not obtained in this session (IEEE Xplore paywalled/anti-bot, no OA mirror found). Getting it — via a university library proxy, interlibrary loan, or a found open mirror — would directly resolve the Part 2 flag and let the "does it really discuss controller/weight adaptation in the NN sense" question be answered from the primary text itself rather than from secondary description.
Schmidhuber's fuller "Annotated History of Modern AI and Deep Learning" (arXiv:2212.11279, 2022) likely has a more elaborated discussion of Dreyfus (1973) than the 2015 overview paper, but its PDF exceeded this session's fetch tooling (no ar5iv/HTML rendering found, arxiv.org/html/2212.11279 returned 404, and the PDF's compressed streams were not machine-readable via the available tools). Worth a dedicated follow-up capture once better PDF tooling access is available.
Werbos's own writing (e.g., The Roots of Backpropagation, 1994, or his 1974 Harvard thesis) was not directly checked in this session for whether he cites Dreyfus's 1973 paper specifically (as opposed to the 1962 paper, which is well-attested). This would be the other half of testing whether 1973 functioned as a causally-transmitted intermediate step (something later lineage members actually read and cited) versus a retrospectively-grouped parallel work.

· web-research batch run, 2026-07-07 · raw markdown