ABSTRACT
The first part of this paper is an introduction to the Machine Learning (ML) methodologies stemming from the Artificial Intelligence (AI) approach. The second part describes a number of possible characteristic features of ML programs. We present a questionnaire as it was submitted to European researchers, then summarize their answers in tables, and conclude with a quick analysis of the answers. The questions we have been asking were chosen to mark the Artificial Intelligence (AI) approach to ML, as illustrated for instance in [Michalski et al. 1983, 1986, Kodratoff 1988]. It follows that some of the systems that perform learning essentially by adjusting coefficients may find it difficult to fit into our schemes. On the contrary, all AI oriented programs should fit in, unless we have made mistakes with our description.
Artificial as it looks, this effort of characterization addresses two deep concerns. The first one is a clarification of the terms used, and sometimes abused, in the ML community. The second is to promote communication in the sub-field. In most cases, learning systems do learn, but it is neither clear what nor how they learn. Setting bench marks for ML is a necessity the community should be more aware of.
Some of the systems known as knowledge acquisition systems perform some learning as well. The last section is devoted to this kind of program.
