Machine Learning

Machine learning is primarily devoted towards building computer programs able to construct new knowledge or to improve already possessed knowledge by using input information; much of this research employs heuristic approaches to learning rather than algorithmic ones. The most active research area in recent years [284] has continued to be symbolic empirical learning (SEL). This area is concerned with creating and/or modifying general symbolic descriptions, whose structure is unknown

a priori

. The most common topic in SEL is developing concept descriptions from concept examples [234], [284]. In particular, the problems in attribute-based spaces are of practical importance: in many such domains it is relatively easy to come up with a set of example events, on the other hand it is quite difficult to formulate hypotheses. The goal of a system implementing this kind of supervised learning is:

Given the initial set of example events and their membership in concepts, produce classification rules for the concepts present in the input set.

Depending on the output language, we can divide all approaches to automatic knowledge acquisition into two categories: symbolic and non-symbolic. Non-symbolic systems do not represent knowledge explicitly. For example, in statistical models knowledge is represented as a set of examples together with some statistics on them; in a connectionist model, knowledge is distributed among network connections [335]. On the other hand, symbolic systems produce and maintain explicit knowledge in a high-level descriptive language. The best known examples of this category of system are AQ and ID families [281] and [314].