Optimizing Mixed Fuzzy-Rule Formation by Controlled Evolutionary Strategy Matthias Lermer, Hendrik Kuijs, Christoph Reich Institute for Cloud Computing and IT Security Furtwangen University of Applied Science Furtwangen, Germany Email: {matthias.lermer, hendrik.kuijs, christoph.reich}@hs-furtwangen.de Abstract—Machine learning algorithms are heavily applied to address many challenges in various fields. This paper specifically takes a look at use cases from the health sector, as well as the industry 4.0 sector. In both cases, the knowledge about the clas- sification process is as important as the classification itself. One current problem is the disregard of expert knowledge provided by adept human beings. In practice, it is possible and also feasible to learn similar knowledge with machine learning algorithms like artificial neural networks (ANNs) or support vector machines (SVMs). However, time and money could be saved if this expert knowledge was used directly. Right now, this is only possible with more transparent algorithms like rule-based systems or decision trees, where knowledge can be incorporated relatively easily. The approach of this paper shows that rules generated by a mixed fuzzy-rule formation algorithm can be optimized by applying a controlled evolutionary strategy while maintaining the interpretability of the decision-making process. The evaluation is performed by executing the evolutionary strategy proposed in this paper on data from two different industries. Keywords–Evolutionary Strategy; Optimization; Fuzzy Logic; Decision support systems; Industry 4.0. I. I NTRODUCTION Nowadays, there is a trend towards using deep learning algorithms, e.g., Deep Neural Networks (DNN), for almost any kind of Machine Learning problem [1]. One of the earlier disadvantages, the slow computation with those kind of algorithms, has been overcome successfully with the help of graphics cards and their optimized cores [2]. Still, one of the big remaining problems is the interpretability of the results when using black box algorithms like DNNs [3][4]. There are many recent approaches to make those results more transparent, but those are still in their infancy [5][6][7]. Other Machine Learning algorithms are more transparent, e.g. Rule- based systems or Decision trees and can provide a human understandable explanation. In practice however, this trans- parency often comes with the price of worse prediction results. The approach depends on the use case or the Machine Learning problem itself. Is it more important to absolutely get the best result possible? Or can a weaker result be tolerated if explanations and knowledge about the results origins can be acquired? In case of the two different scenarios evaluated in this paper, the transparent way to the result is as important as the outcome itself. The remainder of the paper is organized as follows: Sec- tion II provides an overview about related work. Section III describes the genetic adaptation of the Mixed Fuzzy-Rule Formation. In Sections IV and V, the evaluations based on two different Use Cases are conducted. Section VI completes the paper by drawing a conclusion and suggesting future work. II. RELATED WORK Elsayed et al. [8] combine fuzzy rules and evolutionary algorithms, albeit in a different way than in our approach. In their solution, two algorithms cooperate by using fuzzy rules with complementary characteristics. This results in a higher success rate when applied on different data sets with different optimization problems. Their approach is especially interesting as it can be used to further optimize the method proposed in this paper. Schaer et al. have shown that the adjustment of established fuzzy rules and fuzzy set functions can lead to better results [9]. Their work was evaluated within an autonomous car racing competition where they could improve the previous score by 0.5 %. The adjustments and optimizations of the fuzzy components were mainly the product of simulation experiments. In the conclusion, they are mentioning that there are plans to use genetic algorithms for the adjustments which is similar to the evolutionary strategy approach proposed in this paper. Jariyatantiwait and Yen [10] follow the special approach of Differential Evolution (DE). They apply their modification on the ZDT (Zitzler, Deb and Thiele) and DTLZ (Deb, Thiele, Laumanns and Zitzler) test suits [11], which are used for evaluating the optimization of algorithms and map the optimization directly to fuzzy rules. Those rules adapt certain control parameters during the evolution process. Examples are the degree of greediness and exploration. They successfully show that performance metrics can be combined with human understandable knowledge in the form of fuzzy rules. The work conducted in this paper takes a similar approach, but tries to combine classification tasks themselves with fuzzy rules while control parameters like the degree of exploration are defined by hand. Alcal´ a-Fdez et al. [12] show that their modification of a evolutionary fuzzy-rule based system leads to an improved performance within monotonic classification problems. In con- trast to this paper, the authors used genetic algorithms and concentrated on adjusting crossover mechanisms, including customised incest prevention and restarting processes while the mutation mechanism was kept relatively simple by hardcoding the mutation rate. The works from various other authors in this section show that evolutionary strategies within classification problems hold a high value, given the good results and the preserved interpretability by humans. This can be observed for many more use cases, e.g., financial market [13] [14], medicine [15], computer science [16], etc. and reinforce the choice to take a deeper look at the two uses cases of this paper. However, a direct comparison to other works with different use cases is 69 Copyright (c) IARIA, 2018. ISBN: 978-1-61208-631-6 ALLDATA 2018 : The Fourth International Conference on Big Data, Small Data, Linked Data and Open Data