Autonomous Learning Systems: From Data Streams to Knowledge in Real-time

9

Autonomous Learning Controllers

The idea for self-learning controllers is not new and is, perhaps, at the origins of the very idea of self-learning and self-organising systems taking its roots from the very strong, at that time, Moscow Institute of Control Problems (IPU), (IPU was also the work place of Vladimir Vapnik, the ‘father’ of SUM) and mainly related to the works of Tsypkin (1968). This gave the seed for the powerful modern adaptive control theory (Ljung, 1987; Astroem and Wittenmark, 1989). It was and still is, however, mostly valid for linear systems (Kailath et al., 2000) or so-called Hammerstein-type quadratic models and concerns parameter tuning rather than system-structure adaptation.

Later, Procyk and Mamdani (1979) proposed their self-organising fuzzy logic controller (FLC) that was, however, confined to a fixed-size look-up-table, thus, the structure adaptation was very limited and related to the choice of predefined fuzzy sets. Narendra and Parthasarathy (1990) extended the adaptive control systems theory to NN-based multimodel systems, but this was again limited to the case of a fixed system structure and concerned parameter tuning only. Psaltis, Sideris and Yamamura (1988) proposed to model the inverse plant dynamic in an adaptive control scheme using an offline-trained NN and use this to derive controller that would get the performance (output) as desired assuming the plant dynamic is perfectly modelled.

Angelov (2002, 2004b) and Angleov and Buswell (2001), ...