Interpretability Issues in EFS

Abstract

All the previous chapters were dealing with concepts, methodologies and aspects in precise evolving fuzzy modelling, i.e. fuzzy models were automatically evolved based on incoming data streams in order to keep the models up-to-date with the primary goal to obtain on-line predictions from the models with high precision and quality. No attention was paid to interpretability aspects in the evolved fuzzy models, which are one of the key motivations for the choice of fuzzy systems architecture in an (industrial) on-line learning scenario (as discussed in Chapter 1). This deficit is compensated in this chapter, which will deal with the following aspects in EFS:

• Complexity reduction (Section 6.2)

• Towards interpretable EFS (Section 6) which is divided into

  • Linguistic interpretability (Section 6.3.1) and

  • Visual interpretability aspects (Section 6.3.2)

• Reliability aspects in EFS (Section 6.4)

The first issue (complexity reduction) will describe techniques and ideas on how to perform incremental on-line reduction of complexities, i.e. reducing the number of fuzzy sets and rules in evolved fuzzy models, while still keeping the precision of the models high. The outcome of such techniques can be objectively measured in terms of complexities of the reduced models and their predictive performance. The second issue goes significantly beyond pure complexity reduction steps by describing aspects and methodologies for guiding the evolved fuzzy models to more interpretable power, both in linguistic and in visual form. Of course, it is always a matter of subjective taste, whether one or another fuzzy model is more interpretable, but some principal guidelines for achieving more interpretable models (partially motivated from cognitive sciences) can be given. The fourth issue (reliability) embraces the concept of uncertainty in the predictions of fuzzy models, which can be seen as valuable additional information to a human being operating at the machine learning system.