Online Feature-Generation of Code Fragments for XCS to Guide Feature Construction
Trung B. Nguyen, Will Neil Browne, Mengjie Zhang · 2019
Code Fragments (CFs) are a new representation for classifier conditions in Learning Classifier Systems (LCSs). CFs are Genetic Programming-like trees that use functions as internal nodes, and data input or previously learned CFs as leaf nodes for feature construction. The XCSCFC system used CFs in rule conditions of XCS, an accuracy-based Michigan-style LCS, to transfer knowledge and thus solve large-scale problems. However, the trade-off for the richness and flexibility that allows CFs to compactly describe decision boundaries results in an undesired increase in the search space of solutions. Therefore, this paper proposes a novel model extension for Online Feature-generation (OF), which enables evolving features (CFs) through an online observed list of CFs. This extension enables a method of estimating the worth of CFs to identifying the patterns in the problem in order to construct applicable high-level features. The experiments show that the XCS with OF (XOF) can solve the benchmark problems in fewer generations compared with XCSCFC in non-transfer learning scenarios. The novel search of CFs successfully built high-level features, which show the rules produced by XOF to be more generalised than previously possible. Consequently, the final solutions contain fewer rules to solve problems as they encode more compact and comprehensive decision boundaries.