Problem-specific Rule Extraction for Better Performance of Evolutionary Algorithms

Evolutionary algorithms (EAs) are ubiquitous to many environmental issues. Typically, most evolutionary techniques use just a small part of the solutions evaluated at a time. Thus, many of the solutions evaluated during the search process are “forgotten” after one generation, and combined experience of several generations is typically not well exploited. Data mining (DM) techniques can enable deeper insight into the many “good” solutions that have been just simply glimpsed and have been rapidly disregarded because they were dominated by better solutions during an ephemeral moment in the evolution process. Based on a database obtained by suitably recording certain of those disregarded solutions, data mining techniques can help better understand and describe how a system could react or behave after the introduction of changes. This paper proposes applying DM techniques to the set of solutions evaluated after several generations of a single run of an EA in order to extract rules intended initially to be used by the following generations.