Machine learning (ML) played a central role in artificial intelligence at its very beginning. Although the AI limelight has wandered away from machine learning in the advent of other significant developments, such as problem solving, theorem proving, planning, natural language processing, robotics, and expert systems, ML has returned cast in multiple guises, playing increasingly more significant roles. For instance, early work in linear perceptrons faded away in light of theoretical limitations, but resurged this decade with much fanfare as connectionist networks with hidden units able to compute and learn nonlinear functions. In the interim, many symbolic machine learning paradigms flourished, and several have evolved into powerful computational methods, including inductive concept acquisition, classifier systems, and explanationbased learning. Today, there are many active research projects spanning the gamut of machine learning methods, several focusing on the theory of learning and others on improving problem solving performance in complex domains. In the 1980s, the field of machine learning has re-emerged one of the major areas of artificial intelligence, with an annual ML conference, an established 1,000subscriber journal, dozens of books, and ample representation in all major AI conferences.
Perhaps the tenacity of ML researchers in light of the undisputed difficulty of their ultimate objectives, and in light of early disappointments, is best explained by the very nature of the learning process. The ability to learn, to adapt, to modify behavior is an inalienable component of human intelligence. How can we build truly artificially intelligent machines that are not capable of self-improvement? Can an expert system be labeled "intelligent," any more than the Encyclopedia Britanica be labeled intelligent, merely because it contains useful knowledge in quantity? An underlying conviction of many ML researchers is that learning is a prerequisite to any form of true intelligence--