Connection Weight Optimization in a Long Short-Term Memory Using Whale Optimization Algorithm

Abstract

Many academics have recently developed an interest in the procedure for learning LSTM, as well as recognized as an example trickiest issue in artificial intelligence. In most cases, local optima stagnation and slow convergence are the main drawbacks of the standard training procedure. Therefore, it is possible to rely on the stochastic optimization approach to handle these problems. LSTM training has been done using a variety of evolutionary and swarm-based techniques, but the problem of local minima and MSE still exists. Therefore, it was suggested in this study endeavor to optimize the link weights in a long short-term memory using the whale optimization technique. The study compares and presents the results from several datasets in greater depth. (Bladder, Leukemia, Lung, Ovary, Pancreas, and, separately, Prostate cancer) The classification accuracy, convergence rate, and precision score with the highest performance will be determined using MATLAB R2021a. The recommended WOA trainer is compared to GA. Experimental results show that the proposed WOA came out on top for classification accuracy in four instances (bladder, ovary, leukemia, and prostate cancers), whereas GA came out on top in two instances (lung and pancreas cancers). The proposed model outperforms the other algorithm in terms of average convergence speed (MSE) for the datasets related to bladder and ovary cancer, with average MSEs of 0.01888 and 0.006428, respectively. The GA, on the other hand, came in second place, with average MSEs of 0.006984, 0.007946, 0.009627, and 0.008148, respectively. Due to its extensive exploration and avoidance of local optima, the WOA was able to demonstrate results that were superior to those of the other algorithm in terms of convergence. This study further demonstrates that the suggested trainer can successfully train RNN to classify datasets with varying degrees of difficulty.