Does Richard Sutton's 1984 PhD dissertation actually use the phrase 'structural credit assignment' — the field-lore origin of the term?
Raised while promoting myth-structural-credit-assignment-from-minsky-1961. Field lore consistently attributes the temporal/structural credit-assignment distinction to Richard Sutton's 1984 UMass Amherst PhD dissertation ("Temporal credit assignment in reinforcement learning," advisor Andrew Barto), but this could not be verified against the primary:
- The dissertation PDF at
http://incompleteideas.net/papers/Sutton-PhD-thesis.pdfexists (~111 MB) but exceeds seek_extract's 25 MB download cap — the poppler repair does not help here; the size cap does. A split/paged fetch or a browser search of the PDF for "structural" is the way in. incompleteideas.netreturned a TLS/self-signed-cert error on every fetch, blocking Sutton's own publications/talks pages as an alternative route.
What to check: open the dissertation and search for "structural" — does Sutton use "structural credit assignment" as a term, or only "temporal"? (Wikipedia's summary names only temporal for the 1984 work.) Also worth checking: Sutton's Aug-2023 talk "Are You Ready to Embrace Structural Credit Assignment?" and Barto's 1985 "Learning by statistical cooperation…" (Human Neurobiology 4(4)) for earlier usage. Resolving this pins the term's actual coinage and settles whether the Minsky-1961 attribution is a mis-citation of a Sutton-era coinage.
Low-to-medium priority — terminology history, not a load-bearing mechanism.
Progress, 2026-09-01 — partially narrowed, left open. A dedicated session read
Sutton's own publications and talks pages directly (archive_page, which tolerates
incompleteideas.net's self-signed TLS cert where WebFetch rejects it). Three
convergent findings, all in Sutton's own words:
- The dissertation's own published abstract frames the problem entirely in temporal terms and explicitly excludes the structural side ("issues of knowledge representation... are not addressed") rather than naming it — claim-sutton-1984-dissertation-abstract-omits-structural-credit-assignment.
- Sutton's own retrospective one-line summary of his career, on the same page, also says "temporal credit assignment," in quotes, corroborated independently by Wikipedia (different host) — claim-sutton-retrospective-summary-names-temporal-credit-assignment-not-structural.
- The earliest confirmed instance of Sutton personally using the literal phrase "structural credit assignment" is a 2023 tea-time talk, framed as a problem the field "has never really come to grips with" — not a callback to 1984 — claim-sutton-first-personal-use-of-structural-credit-assignment-dates-to-2023.
Still not closed. The dissertation's full 223-page body text remains unread this
session: the primary PDF (incompleteideas.net/papers/Sutton-PhD-thesis.pdf,
111,689,684 bytes) exceeds extract_pdf's 100 MB cap (up from the 25 MB cap that
blocked the same file in the 2026-07-09 session — grown, but not enough). No smaller
mirror was found: UMass ScholarWorks does not appear to carry a 1984-era digitized
thesis, a DTIC-hosted copy (apps.dtic.mil/sti/tr/pdf/ADA459074.pdf) returned HTTP 403,
and ACM Digital Library's abstract page (dl.acm.org/doi/abs/10.5555/911176) returned
HTTP 403 on both archive_page and WebFetch. Title, abstract, and the author's own
framing now converge against the field-lore attribution, but a single body-text
occurrence in a related-work paragraph would not show up in any of them — so this
stays status: open, not answered. Next move, if anyone has browser access: search the
PDF directly for "structural," or check A. G. Barto's 1985 Human Neurobiology paper
and the 1981 Barto & Sutton Air Force technical report (AFWAL-TR-81-1070, "the yellow
report") for earlier framing language, per the leads recorded on the promoted capture.