A Study on the Fitness of GA for Improving SP Decoding Performance

We have studied low-density parity-check (LDPC) coding and iterative decoding methods for shingled magnetic recording (SMR) in high-capacity hard disk drives (HDDs), which is an areal recording density of 4 Tbits/inch<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>. Previously, we applied a neural network to evaluate the log-likelihood ratio (LLR) associated with row operations in the sum-product (SP) decoder for LDPC codes. Then, we emphasized the LLR considering the influence of noise depending on the recording pattern by providing the LLRs for the decoding target and its adjacent bits to the neural network in SP decoding. Furthermore, we explored the optimal parameters to emphasize the LLR by applying the genetic algorithm (GA). In this study, to explore more optimal parameters, we propose a fitness function to improve selection accuracy and increase the number of eligible LLRs. Then, we aim to improve the performance of SP decoding by the effect of GA. As a result, by applying the proposed fitness function to the GA, it is possible to maintain high selection accuracy and to increase the probability that LLRs are emphasized in SP decoding. Therefore, it achieves error-free performance with fewer iterations of turbo equalization compared to the case of a conventional fitness function.