Multi-objective optimization of extreme learning machine using physical programming

Feedforward neural networks have been widely used in various fields, such as disease detection, object tracking, and nonlinear prediction. The performance objectives which are required in different practical problems are also different. It is an issue that how to select the neural network structure to meet the requirement of the designer. This paper presents an algorithm called physical programming (PP) which optimizes multiple performance objectives of networks by selecting the number of hidden nodes and the activation function for extreme learning machine (ELM). In PP, designer's expectations for each objective are divided: unacceptable, highly undesirable, undesirable, tolerable, desirable, and highly desirable, of which the value ranges are decided based on the actual situation and designer's preferences. And then the designer seeks the optimized network structure by genetic algorithm (GA). The simulation result shows that the optimized ELM realizes multi-objective optimization.