ch-oracle-program
2 The ORACLE Program
Learning a compositional world from interaction
ORACLE begins one level before decision making. An Online Rational Agent for Categorical Learning interacts with an unknown categorical environment and constructs a predictive model of its compositional structure. It knows the doctrine of categories, functors, natural transformations, and universal properties, but this is not its only prior. Chapter 0 suggested a stronger developmental hypothesis: the learner begins with a small categorical grammar of objects, magnitude, space, shape, agency, and social relation. It does not know which particular world realizes that grammar, or how its parts are composed in the world it inhabits.
This is a structural analogue of system identification. A predictive-state representation need not supply an optimal policy; it supplies a sufficient model for predicting the consequences of tests. Likewise, ORACLE first asks which compositions, equations, factorizations, extensions, and universal constructions can be predicted from an expanding interaction history. Action selection is a subsequent specialization.
The primitive learning problem is not to optimize inside a fixed category. It is to identify, up to the distinctions exposed by admissible probes, the core-structured category in which later prediction and decision problems are posed.