摘要：This research aims to propose a version of Fast Genetic Algorithm (FGA), namely Fitness Value Memoization Genetic Algorithm (FVMGA). FVMGA uses the concept of memoization to cache the fitness value of chromosomes that have already been calculated before. It allows FVMGA to bypass unnecessary computation for redundant chromosome configurations, which is especially important when we use expensive fitness functions. For benchmarking purposes, the proposed FVMGA was compared to a traditional GA in the use case of optimizing Long Short-Term Memory (LSTM) hyperparameters for time-series forecasting. Four hyperparameters were being optimized in this study with a total of 38,000 possible combinations. However, the number was drastically reduced to 1,000 with the use of GA. The final results showed that FVMGA was able to compute up to 291% faster than traditional GA while maintaining the quality of the produced models.
