JeffShrager@sri-unix.UUCP (09/22/83)
, Letter Spirit (perception of the style of letters), Jumbo (reshuffling of parts to make "well-chunked" wholes), and Deep Sea (analogical perception). These tightly related projects share a central philosophy: that cognition (mentality) cannot be programmed explicitly but must emerge "epiphenomenally", i.e., as a consequence of the nondeterministic interaction of many independent "subcognitive" pieces. Thus the overall "mentality" of such a system is not directly programmed; rather, it EMERGES as an observable (but onnprogrammed) phenomenon -- a statistical consequence of many tiny semi-cooperating (and of course programmed) pieces. My projects all involve certain notions under development, such as: -- "activation level": a measure of the estimated relevance of a given Platonic concept at a given time; -- "happiness": a measure of how easy it is to accomodate a structure and its currently accepted Platonic class to each other; -- "nondeterministic terraced scan": a method of homing in to the best category to which to assign something; -- "semanticity": the measure of how abstractly rooted (intensional) a perception is; -- "slippability": the ease of mutability of intensional representational structures into "semantically close" structures; -- "system temprature": a number measuring how chaotically active the whole system is. This strategy for AI is permeated by probabilistic or statistical ideas. The main idea is that things need not happen in any fixed order; in fact, that chaos is often the best path to follow in building up order. One puts faith in the reliability of statistics: a sensible, coherent total behavior will emerge when there are enouh small independent events being influenced by high-level parameters such as temperature, activation levels, happinesses. A challange is to develop ways such a system can watch its own activities and use those observations ot evaluate its own progress, to detect and pull itself out of ruts it chances to fall into, and to guide itself toward a satisfying outcome. ... Prerequisits: an ability to program well, preferably in Lisp, and an interest in philosophy of mind and artificial intelligence.