Application of Particle Swarm Optimization for Tuning the SVR Parameters

the prediction of finger pinch force via surface electromyography (sEMG) signals is important in bionic control area. The purpose of this paper was to study how to improve the prediction accuracy while using support vector regression (SVR) to predict the pinch force. Four healthy subjects performed constant-posture force-varying pinch operations. The sEMG signal was acquired using two electrodes while the force signal was recorded by a JR3 sensor. The time domain feature of sEMG and the force signal were then applied as the input of the SVR model. In order to improve the prediction accuracy, the parameters of SVR model were optimized by applying particle swarm optimization (PSO) algorithm. The relative mean square error (RMSE), correlation coefficients (CC), and mean average error (MAE) were calculated as the criteria. The results show that the predicted force is close to the real pinch force by SVR modeling technique. The RMSE results are below 8% and the CC results are above 96% with 4 subjects. Compared with the grid search (GS) method, the PSO-SVR achieves a tradeoff between the accuracy and the computational costs with different kinds of training data.