Comparison of MLP-BPNN and MLP-PSO for Automatic Essay Grading System for Japanese Language Exam

Abstract

In this paper, a study was conducted for a hybrid model for Multilayer Perceptron (MLP) with Particle Swarm optimization (PSO). The PSO was used to replace the Backpropagation method for the weight optimization. The comparison was conducted between MLP-BPNN and MLPPSO for an automated essay grading system for Japanese language exam. The MLP-PSO model achieved a more accurate but less stable result. The MLP-PSO model with 10 particles trained for 15 steps achieves the best result out of the two MLP-PSO models tested, with an average 8.48% error for the grade population. Compared to the MLP-PSO model, it was discovered that MLP-BPNN with Adam optimizer achieves better overall performance and results concerning both the accuracy and the stability of the model.