Good versus bad knowledge: Ontology guided evolutionary algorithms Hayden Wimmer a,⇑ , Roy Rada b a Georgia Southern University, Department of Information Technology, P.O. Box 8150, Statesboro, GA 30460, United States b University of Maryland Baltimore County, Department of Information Systems, 1000 Hilltop Circle, Baltimore, MD 21250, United States article info Article history: Available online 28 May 2015 Keywords: Evolutionary algorithms Knowledge guided Genetic algorithm Ontology Decision trees abstract Good knowledge would be expected to help a knowledge-based algorithm more than bad knowledge. In this research, the precise effect of good versus bad knowledge on evolutionary algorithms is explored. The testable hypothesis of this paper is that good knowledge will have a significant effect on the evolutionary mutation process, whereas bad knowledge will have no significant effect. A knowledge-guided evolution- ary algorithm is developed where ontologies, representing knowledge, are applied to the mutation process. Bad knowledge is represented as a randomly generated ontology, while good knowledge is represented by ontologies constructed with domain knowledge and following a formal ontology develop- ment process. Decision trees are evolved to solve a classification problem. Fitness is classification accu- racy. The experiment is replicated over 2 data-sets from different domains with one being time-series, financial data and the other being wine data. As hypothesized, poorly constructed, or bad knowledge, has no effect while good knowledge is shown to have a significant effect. Bad knowledge, being random in character in these experiments, has understandably no impact on an already random mutation process. However, employing knowledge to guide the mutation process significantly constrains the traversal of the search space. Employing knowledge in an evolutionary algorithm has the potential to increase the efficiency and accuracy of evolutionary algorithms. Ó 2015 Elsevier Ltd. All rights reserved. 1. Introduction Employing codified domain knowledge to guide the evolution- ary mutation process of an evolutionary algorithm (EA) has the potential to increase the efficiency of the EA. The knowledge guides the EA by mutating genes within an organism to genes which are semantically similar. A genetic algorithm (GA) is an EA inspired by natural evolution and capable of locating an optimal solution in a complex landscape (Zarifia, Ghalehjogh, & Baradaran-nia, 2015). GAs have been applied to complex optimization problems, such as logistics scheduling (Chang, Wu, Lee, & Shen, 2014). Combining a GA with another algorithm to improve performance creates a hybrid GA, with one example being a GA for parameter optimization of a support vector machine (Chou, Cheng, Wu, & Pham, 2014). Directed GAs seek to direct the mutation operation to increase the efficiency and performance of a GA (Kuo & Lin, 2013). Employing knowledge to guide the mutation process is another option to improve performance of a GA. This research explores the effect of good versus bad knowledge guiding the evolutionary process of a GA. Ontologies are a shared conceptualization of a domain (Gruber, 1993). Ontologies are used for communication between humans, between human and machine, or between machines, as well as for computational inference and knowledge reuse (Gruninger & Lee, 2002). Good, or well-constructed, knowledge is constructed following formal ontology development processes coupled with domain knowledge. Bad, or poorly-constructed knowledge is defined here as an ontology constructed in a random fashion. Knowledge has been used to improve genetic algorithms by opti- mizing the feature subset selection for input to a GA (Wendt, Cortés, & Margalef, 2010; Yang & Honavar, 1998). Case-based rea- soning has been paired with GAs to find an optimal solution (Huang, Huang, & Chen, 2007). Heuristics and GAs have been paired by using GAs to extract heuristics from data (Gordini, 2014) as well as exploiting heuristics to guide the mutation pro- cess of a GA (Johns, Keedwell, & Savic, 2014; Wimmer & Rada, 2013). The aforementioned research indicates the potential for knowledge guided evolutionary algorithms. Evolutionary algorithms, specifically genetic algorithms, are stochastic in nature and therefore unpredictable and uncontrolled. Controlling the mutation of an EA via constraining the search oper- ation may lead to improved performance via improved fitness or reaching an optimum in fewer generations. Knowledge may be http://dx.doi.org/10.1016/j.eswa.2015.04.064 0957-4174/Ó 2015 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. E-mail addresses: hwimmer@georgiasouthern.edu (H. Wimmer), rada@umbc. edu (R. Rada). Expert Systems with Applications 42 (2015) 8039–8051 Contents lists available at ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa