|
Abstract:
|
Aiming at building an LCS with reinforcement process consistent with the gradient descent update of RL while utilizing XCS's accuracy-based rule discovery process, this paper presents \textit{Dual-structured Classifier System} (DCS), which processes the gradient descent based update and XCS style update in parallel. DCS is evaluated on the popular test-bed problems for LCSs with three types of bucket-brigade algorithms, Q-bucket-brigade, implicit-bucket-brigade, and residual-bucket-brigade each combined with XCS's original bucket-brigade. Empirical results are provided that shows DCS performing better than ZCS with the same optimal level of XCS, while the consistency with RL's gradient descent update enables to apply the results of theoretical analyses in RL field including rigorous conditions for the convergence of DCS's reinforcement process.
|
