Learning concepts from data

摘要

Current data-mining practice employs relatively low-level machine learning algorithms—statistical, neural-net, genetic, decision-tree, etc.—to trawl large data-sets for new classifiers. Usefulness of classifiers is then assessed according to accuracy in classifying new data, e.g. for stockmarket prediction or for assigning causes in medical diagnosis. User-transparency is often a secondary consideration. But in some applications the priorities are reversed. (1) In the “designer drug” industry, predictive classifiers are more valuable if they also provide chemists with new leads, while minimizing the frequency of the false positives. Screening fossil fuel emissions for hazardous components, on the other hand, requires minimization of false negatives. The high user-transparencies required in both cases are obtainable only with relational descriptions and rule languages, going beyond attribute-value representations. (2) In the simulator training of aircraft pilots, and of operators of other real-time devices, predictive control laws can provide a basis for machine-generated real-time advice to trainees. Study of human simulator-based control learning indicates a subgoal structure. It also suggests that both acquisition by the learner of new subgoals, and reorganization of perceived causal links among problem elements are involved in real-time skill learning. Design implications for automated coaches are worthy of consideration.