Does Dreyfus's 1973 'time-lag optimal control' paper actually contain the controllers-as-weights backprop step Schmidhuber attributes to it?

This capture follows up the vault's open question question-read-dreyfus-1973-time-lag-optimal-control-primary and extends claim-dreyfus-1973-lineage-link-uncorroborated. The paper in question: Stuart E. Dreyfus, "The computational solution of optimal control problems with time lag," IEEE Transactions on Automatic Control 18(4):383–385 (1973), DOI 10.1109/TAC.1973.1100330. Genuine effort was made to locate the full text (IEEE Xplore, arXiv, gwern.net, Semantic Scholar, Unpaywall, CORE.ac.uk, base-search.net, CiteSeerX, ResearchGate, DTIC, UC Berkeley IEOR departmental repository) — none carries an open copy. The paper stays paywalled and unread by this research chain, so the core question cannot be confirmed or denied — see Claim 4.

Claim: Schmidhuber's own writing is the sole source of the "controllers-as-weights" framing, in his own words

Jürgen Schmidhuber's peer-reviewed history of deep learning states, verbatim: "Efficient BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973)." This is confirmed independently on two of his own venues with identical wording: the arXiv preprint (fetched and abstract/metadata verified directly) and his personal page "Who Invented Backpropagation?" (fetched directly, verbatim match): "BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973)." The reference-list entry on the same personal page cites the paper's correct bibliographic details. This is a technical-mechanism claim about what Schmidhuber asserts, not about what the 1973 paper actually contains — it is sourced directly to Schmidhuber's own primary statements, clearing the Tier 1–2 floor for mechanism claims.

• source_url: https://arxiv.org/abs/1404.7828 (arXiv preprint; published as Schmidhuber 2015, Neural Networks 61:85–117, p. 11 §5.5 per subagent direct read of the PDF)
• source_url (personal site, independently WebFetched and verified this session): https://people.idsia.ch/~juergen/who-invented-backpropagation.html
• source_author: Jürgen Schmidhuber
• source_date: 2014 (arXiv) / 2015 (Neural Networks); personal page undated, standing content
• source_quote: "BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973)."
• source_tier: 1 (arXiv preprint of a paper subsequently peer-reviewed in Neural Networks; personal site is Tier 2 on its own but corroborates the peer-reviewed wording exactly)

Claim: No open-access copy of Dreyfus 1973 exists anywhere checked; it remains fully paywalled

Bibliographic and open-access-status checks converge: Semantic Scholar's record for the paper shows an empty openAccessPdf field, and Unpaywall returns is_oa: false / best_oa_location: null. No copy was found on gwern.net (which hosts other Dreyfus papers, e.g. the 1990 JGCD note, but not this one), CiteSeerX, ResearchGate, DTIC, CORE, or the UC Berkeley IEOR departmental site that hosts the open Mizutani–Dreyfus–Nishio 2000 PDF. The IEEE Xplore record itself (https://ieeexplore.ieee.org/document/1100330/) resolves as a URL but renders no readable text through available fetch tools — consistent with a standard publisher paywall, not with anything adversarial. This is a historical/access claim (uncontested, non-quantitative) resting on Tier 1–2 aggregator confirmation (Semantic Scholar API, Unpaywall), which clears the floor for this kind of claim.

• source_url: https://www.semanticscholar.org/paper/0cb6bdf8c17e4cafb01e7c523abffda2f76858fb
• source_author: Semantic Scholar (aggregator record for Dreyfus, S.E. 1973)
• source_date: accessed 2026-07-19
• source_quote: "openAccessPdf: null" (API field, paraphrase of structured data — no prose quote exists to lift verbatim)
• source_tier: 2 (aggregator of primary bibliographic/OA-status metadata, not a claim about content)

Claim: Dreyfus's own two retrospectives never cite the 1973 paper and explicitly credit 1962 instead — corroborating evidence beyond what's already in claim-dreyfus-1973-lineage-link-uncorroborated

This is not a new finding — claim-dreyfus-1973-lineage-link-uncorroborated already establishes that the 1973 paper is absent from both of Dreyfus's retrospective reference lists. This capture adds one directly-verified data point that sharpens the same argument-from-silence (observation-bjork-dreyfus-same-argument-from-silence-graded-differently): the closing sentence of Dreyfus's 1990 note names the origin explicitly, in his own words: "The recursive derivation of these formulas using the chain rule, commonly seen in the neural-network literature, was first used for optimal-control problems by Dreyfus" — a sentence that (per the paper's own 7-entry reference list) points to the 1962 JMAA paper, not 1973. The 2000 Mizutani–Dreyfus–Nishio paper's 14-entry reference list is likewise 1973-free. Both PDFs are the same primaries the existing note already rates Tier 1.

• source_url: https://gwern.net/doc/ai/nn/1990-dreyfus.pdf
• source_author: Stuart E. Dreyfus
• source_date: 1990
• source_venue: Journal of Guidance, Control, and Dynamics 13(5):926–928
• source_quote: "The recursive derivation of these formulas using the chain rule, commonly seen in the neural-network literature, was first used for optimal-control problems by Dreyfus."
• source_tier: 1
• companion source_url: https://ieor.berkeley.edu/wp-content/uploads/2019/03/ijcnn2k.pdf (Mizutani, Dreyfus & Nishio, IJCNN 2000) — confirmed no "1973" or "time lag" string anywhere in the text, reference list 14 entries, none is the 1973 paper.

Claim: The core question is unresolved — the only text of the 1973 paper obtainable is a triangulated abstract with no explicit weights/controller language, and the paper's body remains unread

The closest available approximation to the paper's content is its abstract, which does not appear on a directly-fetchable page (IEEE Xplore's page did not render) but was cross-triangulated as identical wording across Semantic Scholar's metadata and independent search-engine snippets: "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." Nothing in this wording explicitly frames control parameters as "weights" the way Schmidhuber's gloss does — but the abstract is not the body, and the body was not read by anyone in this research chain. No independent historian-of-science source was found that has read and describes the 1973 paper's actual content (Griewank 2012's "Who Invented the Reverse Mode of Differentiation?" is a plausible candidate but could not be fetched this session — see Further leads). Wikipedia's "Backpropagation" article states "In 1973, he adapted parameters of controllers in proportion to error gradients," sourced to the same 1973 citation — this is citation-chained from Schmidhuber, not independent verification, and per this vault's sourcing floor is a Tier 4 pointer, not evidence.

Verdict: [unverified — could not confirm or deny after search]. The "controllers-as-weights" framing is confirmed to be Schmidhuber's own characterization (Claim 1), the paper itself is unreachable in full (Claim 2), and Dreyfus's own retrospectives point to 1962 rather than 1973 as his relevant contribution (Claim 3) — but whether the 1973 paper's body actually contains a controllers-as-weights/backprop-style step, or whether Schmidhuber's gloss over-reads a generic history-dependent gradient procedure, remains genuinely open pending access to the full text.

• source_quote (abstract, triangulated not primary-fetched): "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."
• source_tier: this abstract text is [unverified-mechanism — needs primary] — it is triangulated via an aggregator (Semantic Scholar metadata) and matching search-engine snippets, not confirmed by a direct fetch of the IEEE Xplore page itself, and is not adequate on its own to settle a technical-mechanism claim per the sourcing floor.

Further leads

• Andreas Griewank, "Who Invented the Reverse Mode of Differentiation?" (Documenta Mathematica, Extra Volume ISMP, 2012, pp. 389–400) plausibly discusses Dreyfus's control-theoretic priority in more depth than Schmidhuber's survey — fetch attempt to http://emis.icm.edu.pl/journals/DMJDMV/vol-ismp/52_griewank-andreas-b.pdf failed with a connection error this session; worth a retry.
• Scholarpedia's "Deep Learning" article (Schmidhuber, author) could not be fetched (empty response, two attempts) — may restate or expand the same 1973 attribution; unverified.
• A 1972 UC Berkeley Operations Research Center technical report (ORC-72-21) appeared in one aggregator snippet as a possible precursor to the 1973 published paper, but could not be located in any repository checked — a possible alternate route to the content if it survives in an institutional archive.
• Yuxi Liu's blog essay "The Backstory of Backpropagation" (yuxi.ml, 2023/2024) covers the optimal-control/adjoint-method history of backprop in some depth but never mentions Dreyfus 1973 or 1962 at all — a blog-tier (not citable) absence, but a second independent silence worth noting alongside Dreyfus's own.
• The IEEE Xplore abstract page for the paper (https://ieeexplore.ieee.org/document/1100330/) resolves as a URL (redirects correctly from the DOI) but renders no parseable text through the tools available this session — a live target for a future session with different fetch tooling or institutional access.

Safety flags

None. No page encountered this session showed addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing. The IEEE Xplore non-render and the two failed fetches (Griewank PDF, Scholarpedia) are ordinary access failures, not adversarial signals — no tls:"unverified" provenance was recorded for any source in this capture.

Entity candidates

• Stuart Dreyfus — person — central actor across the entire backprop-origins cluster (appears in a dozen-plus existing claim-notes) but has no entity page of his own; only his brother entity-hubert-dreyfus does
• Andreas Griewank — person — historian of automatic differentiation whose 2012 "Who Invented the Reverse Mode" survey is repeatedly cited as load-bearing across this cluster but remains unread by this chain
• IEEE Transactions on Automatic Control — concept — the recurring paywalled venue blocking primary verification across multiple notes in this cluster (Dreyfus 1973, and potentially Kelley/Bryson-adjacent papers)