Automatically defined functions for learning classifier systems

Muhammad Iqbal, Mengjie Zhang, Will N. Browne

Abstract: This work introduces automatically defined functions (ADFs) for learning classifier systems (LCS). ADFs had been successfully implemented in genetic programming (GP)for various domain problems such as multiplexer and even-odd parity, but they have never been attempted in LCS research field before. ADFs in GP contract program trees and shorten training times whilst providing resilience to destructive genetic operators. We have implemented ADFs in Wilson's accuracy based LCS, known as XCS [14]. This initial investigation of ADFs in LCS shows that the multiple genotypes to a phenotype issue in feature rich encodings disables the subsumption deletion function. The additional methods and increased search space also leads to much longer training times. This is compensated by the ADFs containing useful knowledge, such as the importance of the address bits in the multiplexer problem. The ADFs also create masks that autonomously subdivide the search space into areas of interest and uniquely, areas of not interest. The next stage of this work is to implement simplification methods and then determine methods by which ADFs can facilitate scaling for more complex problems within the same problem domain.