Karl Johan Åström

Swedish control theorist who, in 1965 while working at IBM's Nordic Laboratory, published "Optimal Control of Markov Processes with Incomplete State Information" (Journal of Mathematical Analysis and Applications, vol. 10, pp. 174–205) — the paper that gave a full mathematical treatment to the problem of choosing optimal actions when the true state of a system can only be inferred from noisy or incomplete observations.

Matters to this vault as the confirmed next link in a chain it had already started tracking on A.A. Fel'dbaum's page as an unread lead: Fel'dbaum posed the problem of a controller that must simultaneously learn a system's dynamics and act on it (dual control, early 1960s); Åström's 1965 paper is cited as reference [1] in Kaelbling, Littman & Cassandra's 1998 Artificial Intelligence paper, the document that carried the resulting mathematical object — the partially observable Markov decision process — into mainstream AI planning research and, sixty-one years later, into how 2026 papers formalize LLM-agent decision-making.

References

• claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin — the 1965 paper's citation lineage into AI

• claim-kaelbling-1998-bibliography-misdates-astrom-1965-paper-1995 — a thirty-year citation-date error on the same reference

• claim-2026-comap-paper-formalizes-llm-agent-as-pomdp — the same POMDP structure in 2026 LLM-agent research

• Capture: 10-inbox/raw/2026-09-02-hop-astrom-pomdp-llm-agents.md

• 2026-09-03: the thirty-year citation-date error was traced to a specific window — present in every independently-hosted copy of the 1998 paper's final typeset edition, absent from the authors' own pre-publication preprint and 1994 conference paper — localizing it to the paper's own production/copy-editing step rather than any later web re-hosting (claim-kaelbling-1998-misdating-originated-in-final-typesetting-not-web-hosting).

• 2026-09-10: the 1965 paper itself was read directly for the first time (extract_pdf, Lund University Publications reprint), rather than only through its citation history. It frames its own problem entirely in control-theory vocabulary and never uses the word "agent" (claim-astrom-1965-paper-uses-control-vocabulary-not-agent-language); its own introduction nonetheless explicitly claims the result generalizes to other fields, offering queuing theory as an example (claim-astrom-1965-introduction-claims-generalization-beyond-control-systems); its second worked example is imported wholesale from Ronald Howard's 1960 "toymaker" business-decision problem (claim-astrom-1965-second-example-is-howards-toymaker-problem); and its own Notes section is the vault's first primary-source confirmation that Åström directly cites A.A. Fel'dbaum, alongside Bellman, Pontryagin, and Kolmogorov, as his control-theory lineage — "Submitted by Richard Bellman" (claim-astrom-1965-notes-section-cites-feldbaum-pontryagin-kolmogorov-lineage).