Learning and control from a recursive viewpoint

The authors consider the problem of learning to control a physical process. They present the methodology, starting with a program, CANDIDE, which has proven successful for different tasks, including car driving, and introduce the resulting theory. More precisely, they shed light on recursivity, a learning principle that yields interesting results and is a very powerful tool for obtaining absolute limitations of algorithmic control and learning. An application to control and another to learning are reported. Both show that from a recursive viewpoint, certainly some problems concerning control or learning are alike