Drescher G. L. (1989) Made-up minds: A constructivist approach to artificial intelligence. PhD thesis. of Electrical Engineering and Computer Science, MIT. Fulltext at https://cepa.info/5784
Made-up minds: A constructivist approach to artificial intelligence. PhD thesis.
of Electrical Engineering and Computer Science, MIT.
Fulltext at https://cepa.info/5784
The schema mechanism is a general learning and concept-building mechanism intended to simulate aspects of Piagetian cognitive development during infancy. A computer program that implements the schema mechanism, MARCSYST I, has replicated several early milestones in the Piagetian infant’s invention of the concept of permanent object. In Piaget’s constructivist theory, an infant first represents the world only in terms of simple sensory and motor elements; initially, there is no concept of persistent, external objects-objects that exist even when not perceived. The infant must construct this concept, working backward from the perceptions that manifest external objects. This conceptual leap is first of a long series of such constructions, extending through adult-level intelligence. The schema mechanism connects to a simulated body in a microworld. The mechanism learns from its experiences by processes of induction, abstraction, and invention. A novel induction technique builds schemas, each of which asserts that a given action, in certain contexts, has particular results; contexts and results are expressed in terms of binary state elements called items. Crucially, the schema mechanism not only thus discovers relations among existing representational elements (actions and items), but also constructs new such elements. For any achievable result, the mechanism can define a new, abstract action, the action of achieving that result. Most important, the mechanism can synthesize new state elements to designate aspects of the world that the existing repertoire of representations fails to express, thus inventing new concepts. The schema mechanism builds schemas, expressing the context-dependent results of actions, by an induction technique called marginal attribution. Discovering the results of actions is complicated by the fact that a particular action typically has different effects on different occasions. Until the corresponding context conditions have been identified, a result is therefore difficult to identify as such, and vice versa. The marginal attribution facility solves this chicken-and-egg problem by identifying a relevant state transition-one which, even if it follows a given action only rarely, is even more rare in the absence of the action. Then, the mechanism searches for conditions under which the relevant result follows more reliably. The schema mechanism can define a new state element to represent the validity conditions of an unreliable schema. For example, suppose a given schema asserts that moving the hand to a certain body-relative position results in a tactile sensation. The schema mechanism defines a new state element to represent whatever unknown condition must hold for the schema to be valid; in this case, the condition is that there be a palpable object at that position. The concept of palpable objects is not built in; rather, this synthetic item itself implements the mechanism’s first approximation to that concept, thus reifying the schema’s validity conditions, treating the schema’s validity as a thing-in-itself. Having defined a new state element, the mechanism begins an open-ended process of finding the element’s verification conditions, which tell about its state.