Stabilizing the Convergence of Online-Learning in Neuro-Fuzzy Systems by an Immune System-inspired Approach

Adaptive learning systems are needed in many control applications. Neuro-fuzzy systems are very useful here because they are universal function approximators and do not require a formal process model. Furthermore, they allow incorporating a priori knowledge to speed up convergence or to treat safety-critical situations. But learning in a closed loop setup may nevertheless lead to a truly chaotic systems behaviour because what is learned has in impact on what is learned next, and so forth. Hence an arbitrary control behaviour may emerge dynamically, thus leading to an allowed, but suboptimal control behaviour or learning result, respectively. The dynamic learning process thus has to be guided, in order to avoid such unwanted system characteristics. This is the aim of the SILKE-approach (System to Immunize Learning Knowledge-based Elements). This paper describes its basic concept and presents some first results.