Applying XCS on time variant problem: Separates thinking from doing

摘要

Extended Classifier System (XCS) has been proved to be a fine classifier for pattern recognition tasks and was adopted as a popular research tool for several active research fields. During the progress of developing XCS, several versions of XCS such as XCS with real value attribute (XCSR), XCS with additional memory (XCSM) and parallel XCS (DXCS), XCS as function approximator (XCSF) have been proposed to meet the needs of real world applications. On the other hand, in the field of biomedical engineering and financial time series forecasting, data gathered is inherently time variant, while both of them are most active research fields nowadays, it would be valuable to gain more insights from how XCS works when encounter with time variant data. Hence, in this study we examined XCS's performance on time variant problem and proposed an alternative version of XCS based on simulating human nature that combing wild guessing on everything and careful reaction together by separating thinking and acting components in the design of XCS. The results showed that the new version XCS (97.11% accuracy rate in average) out performed traditional XCS (77.73% accuracy rate in average), by significance level of p < 0.0001 on time variant 6-multiplexer problem.