Genetic programming based predictions and estimations for the endurance and retention of NAND flash memory devices

The central hypothesis of this thesis is that it is possible to use a supervised machine learning technique, Genetic Programming (GP), to make accurate predictions and estimations regarding the endurance and retention of multi-level cell NAND Flash Memory devices. The retention of storage locations, or blocks, within these devices is the length of time for which they successfully retain their data, while their endurance is the number of times they can be programmed and erased prior to failure. Manufacturers currently place conservative speci cations on their devices since there is no technique available to quickly determine the actual endurance and retention capabilities of blocks within them. An extensive empirical evaluation of a number of MLC NAND Flash devices is completed, identifying features for use with GP, before expressions are evolved to make predictions and estimations regarding the retention and endurance of blocks. The empirical evaluation highlights the large variations in performance between blocks in di erent devices of the same speci cation, and even between blocks within the same device. As well as building a data set for later use with GP, the durations of program and erase operations are identi ed as features with which to make endurance predictions and estimations, while a relationship between block location and endurance is also established. GP is employed to evolve binary classi cation expressions, referred to as retention period classi ers, to predict whether blocks will correctly retain their data for a speci ed length of time. Following this, endurance classi ers are evolved to predict whether blocks will successfully complete a prede ned number of cycles. Finally, symbolic regression expressions are evolved, building on the earlier experiments, to estimate the actual number of cycles each block will complete prior to failure and are referred to as endurance estimators.