A Concept Learning Algorithm with Adaptive Search
Kazumi Saito, Ryohei Nakano · 1996
Abstract This chapter describes a concept learning algorithm called RF4 which adaptively improves its concept learning efficiency. RF4 performs a depth-first search on the basis of five criteria for pruning undesirable formulae and five transformation rules for combining formulae. The attribute search order is determined dynamically by estimating each attribute’s probability of being a component of the concept. In a KRK chess endgame problem, after learning a few sets of training examples, RF4 quickly improved its learning efficiency as well as predictive accuracy. In a set of visual pattern recognition problems called Bongard problems, using primitive knowledge of graphical objects, RF4 solved 41 out of 100 problems efficiently; other learning algorithms, e.g. GOLEM, INDUCE, or FOIL solved much less. After solving the 41 training problems, the average time for RF4 to solve the same set of problems was reduced to about onethird. Through statistical tests, a piece of useful knowledge to improve the problem-solving efficiency was extracted from a set of probabilities used in RF4.