APPLICATION PARTICLE SWARM ALGORITHM TO MINIMIZE THE COST OF CONDUCTING MULTIVARIATE EXPERIMENT
Nikolay D. Koshevoy, A. A. Beliaieva · Radio Electronics Computer Science Control · 2018
Context. The actual problem of obtaining a sequence of experiments in the conduct of a full factor experiment ensuring its minimum costhas been solved.Objective – is to create a method for optimizing multifactor experimental plans using an optimization algorithm for the particle swarm.Method. A method is proposed for constructing an optimal experiment design matrix for the cost of implementation using the particleswarm algorithm. The particle swarm method is based on modeling the behavior of the particle population in the parameter space of theoptimization problem. In the beginning, the number of factors and the cost of the transition for each level of factors are introduced. Then,taking into account the input data, a composite matrix of experiment planning is formed. The particles are scattered randomly across theentire composite experiment design matrix and each particle has a random velocity vector. After that, the particles begin to move along therows and columns of the matrix. At each point where the particle visited, the value of the experiment is calculated. In this case, each particleremembers which (and where) the best value of the cost of the experiment, she personally found and where the point is located, which is the best among all the points that explored the particles. At each iteration, the particles correct their velocity (module and direction) in order to be closer to the best point on the one hand, which she found herself and, at the same time, to approach the point that is currently globallybetter. After a certain number of iterations, the particles are collected near the best point. Then the current coordinate of each particle iscorrected. After this, the cost of the experiment is calculated at each new point, each particle checks whether the new coordinate has becomethe best among all the points where it visited. Then, among all the new points, we check whether we have found a new globally better point,and if found, remember its coordinates and the value of the cost of conducting the experiment in it. Then the gain is calculated in comparisonwith the initial cost of the experiment.Results. The software that implements the proposed method is developed, which was used in carrying out computational experiments tostudy the properties of the method.Conclusions. The conducted experiments confirmed the efficiency of the proposed method and the software that implements it, and alsoallow them to be recommended for application in practice when constructing optimal experimental design matrices.