Learning by theory refinement

Type: kb/types/definition.md · Tags: foundations, self-improving-systems, learning-theory

Learning by theory refinement is the learning paradigm in which a system retains an addressable tentative theory outside its model weights, the theory guides its decisions, the outcomes of that work refine the theory, and the refined theory guides later work. What the system has learned is the change in its later behavior that is attributable to the change in the retained theory.

The KB needs the term because theory refinement names one operation, a single episode of revising a theory, while the research program's claims concern the recurrent loop built on it. Without its own name the loop borrows the operation's name, and claims about one are read as claims about the other. The compound is the technical term.

The loop has four parts, and each is necessary:

  • Retained. The theory survives between decisions as an object that can be inspected and revised.
  • Guides decisions. The theory is on the causal path of a decision: changing or withholding it would change what the system does.
  • Refined by outcomes. Outcomes of the work bear on the theory, and the theory is revised against them part by part. This is theory refinement.
  • Guides later work. Later decisions use the refined state. This closes the loop.

The last three correspond to the claims of the evidence ladder: mediation, empirical contact and theory refinement, and recurrent mediation. For example, a system whose account of which edits need which checks fails on one included snippet, revises the part of the account that failed, and afterwards checks snippets it has never seen fail, has completed one turn of the loop.

The loop resembles Popper's schema for the growth of theories, P1 → TT → EE → P2: a problem, a tentative theory, attempted error elimination, and a new problem. The resemblance is our reading, and it is loose in one respect. In Popper's schema "the result of criticism, or of error-elimination, applied to a tentative theory, is as a rule the emergence of a new problem" (Popper 1966, verbatim), not a revised theory. The loop defined here adds what the schema leaves open: the theory that survives or replaces the criticized one is retained, and it guides the work in which the next problem is met.

Scope

  • Not a success term. A loop that carries a mistaken theory forward is still learning by theory refinement. Whether it improved anything is measured separately.
  • Fixed weights are a study condition. Holding the weights fixed rules out parameter updates as the source of a change, which isolates the loop. It is not part of the definition, and a system may run the loop and adapt its weights as well.
  • Any form, machinery, and subject. The term inherits these from the operation. The theory may be prose, a program, or a mixture; a model or a program may do the refining; and the theory may describe an external subject or the system's own organization, which is the reflective case.
  • The carrier is separate. The system that runs the loop is a theory builder, which is identified by responsibility and lineage and need not learn.
  • Records may be kept alongside. Retaining the records a theory was derived from, as evidence for re-examining it, is compatible with the paradigm. What the definition requires is that a theory is retained and refined, not that nothing else is.
  • What the definition leaves open. It says what the paradigm is, not how to measure it. What evidence establishes each part, and how a later change is attributed to a revision, belongs to the evidence ladder and to the experiments, and some of it is unsettled. A partial loop is reportable at its own strength: a system that shows guidance or one refinement without later reuse has shown that part, not the paradigm.

Exclusions

  • A theory built while reasoning and discarded after the decision. It can guide that decision, but nothing remains to refine.
  • Keeping records or summaries of experience and rebuilding an explanation when one is needed. No explanation is retained as an object to revise.
  • Adapting weights. What is learned is not a theory with parts that can be inspected and revised.
  • A retained theory that is applied and never revised, such as fixed instructions. The system uses a theory; it does not learn by refining one.
  • A stored theory nothing consumes, since a representation matters only through its consumption path.

Misuse Cases

  • Calling a system a learner by theory refinement because it retains prose about its subject or about itself. The term names a loop on the causal path of decisions, not an artifact.
  • Using theory refinement for the loop, or learning by theory refinement for a single revision.
  • Treating fixed weights as part of the definition.
  • Reading the term as a claim that the paradigm works or beats the alternatives. That is the conjecture in learning by theory refinement may improve sample efficiency under structured shifts, and the definition is neutral on it.

Relevant Notes: