Evolving System Structure from Streaming Data
5.1 Defining System Structure Based on Prior Knowledge
Traditionally, the system structure (whichever type it is, e.g. based on probabilistic models, neural networks, fuzzy rule-based systems, polynomial models, etc.) is being predefined and fixed. The choice of the structure (Bayesian, hidden Markov models, number of states, neurons, rules, order of the polynomial, type of distributions, activation functions, membership functions, number of inputs/features, etc.) is usually based on prior knowledge and insight from the problem domain. For this reason, such an approach is problem-specific, expert-dependent (therefore, not suitable for autonomous and online behaviour) and ignores the possible dynamic evolution of the problem at hand. The last factor becomes more and more important nowadays.
The innovative approach on which this book and the author's research in the last decade is based is focused on the development of objective, automatic, adaptive and autonomous methods for system structure identification (in a dynamic context) from the data streams. This approach was called evolving systems a decade ago (Angelov, 2002).
For example, initially, fuzzy systems structure identification was based on the use of prior expert knowledge (Zadeh, 1975; Driankov, Hellendoorn and Reinfrank, 1993). This was a logical step since the introduction of the fuzzy set theory and fuzzy linguistic variable by L. A. Zadeh was seen primarily as a tool and ...