Perbezaan Di Antara Teknik Pengawalan Nilai Awal Pemberat Dalam Algoritma Backpropagation

Standard Backpropagation Algorithm (BP) is a widely used algorithm in training Neural Network that is proven to be very successful in many diverse applications. Despite the general success of this algorithm, there are several drawbacks and limitations which some of them are the existence of local minima, slow rates of convergence, problem in generalization and some of the modification of BP algorithm requires complex and costly calculations at each iteration, which offset their faster rates of convergence. Normally, this algorithm use random technique in order to initialize weight for training purpose. Many researchers had approved that weight initialization give big effect to performance of BP algorithm. Shimodaira (1994) had proposed OIVS (Optimal Initial Value Setting) technique for weight initialization to increase performance of algorithm. Other than that Drago et al (1992) had proposed SCAWI (Statiscally Controlled Activation Weight Initialization) technique and Widrow et. al (1990) had proposes Nguyen-Widrow technique to increase performance of algorithm. This project is investigating the effectiveness of BP algorithm with different initial weight initialization technique. This experiment use three different data set which is Balloon, Iris and Cancer data set where Ballon is for small scale data set, Iris for medium scale and Cancer data set for large scale. The results display the performance of each technique and also demonstrate that Nguyen-Widrow is the best initial weight initialization technique.