Misclassification detection based on conditional VAE for rule evolution in learning classifier system

This paper focuses on the problem of Learning Classifier System (LCS) that is hard to guarantee to generate the "correct" output (i.e., the action in LCS) as the dimension size of data increases (which results in producing the "incorrect" output) and proposes the method that can detect the incorrect output of LCS. For this issue, this paper proposes the Misclassification Detection based on Conditional Variational Auto-Encoder (MD/C) which detects and rejects the incorrect output of LCS through a comparison between the original data and the restored data by CVAE (Conditional Variational Auto-Encoder) based on the output of LCS (as the condition to CVAE). The results of a ten-class classification problem using handwritten digits showed that MD/C properly rejects the incorrect output of LCS and achieves 99.0% correct rate.