Performance Analysis of Self-Organizing Map Method for Wheeled Robot Control System

Abstract

Self-organizing map (SOM) is one of the unsupervised learning techniques and is the alternative control system for wheeled robots in addition to the supervised learning techniques such as backpropagation neural network (BPNN). This research aims to compare the performance between SOM and BPNN control systems for wheeled robot. Direct inverse control (DIC) is utilized as the control system by generating a control signal based on the determined trajectory through the inverse process. The implementation of the DIC system with dynamic input-output mapping requires a modification for the original SOM algorithm. Its modification utilizes the vector-quantized temporal associative memory techniques. The SOM control system shows better performance because it produces a lower error compared to BPNN. It also only requires a constant 131 epochs for each training compared to the Bayesian regularization BPNN control system which requires an average of 253 epochs for each training. These results show that the SOM control system can produce a lower control error with less computational cost compared to the BPNN control system.