Learning and generalisation. With applications to neural networks. (Q1856371)

In this second edition [for a review of the first one, see (1997; Zbl 0928.68061)], according to its preface, two main innovations are to be found: First, the hypothesis of independent and identically distributed samples of the learning algorithm is weakened to a mixing one which is seen to appear in certain Markov processes. Second, applications in systems science via the utilization of randomized algorithms is twofold: a) Many synthesis algorithms which are NP-hard in a deterministic context become P hard when randomized. b) Identification results provide finite-time estimates with the methods of this book while only asymptotic results are obtained in the classical way; this is useful in design when identification and control are combined. Some chapters have been modified to take into account of recent advances.