Multi-Population-Based Algorithm with an Exchange of Training Plans Based on Population Evaluation

Abstract

Population Based Algorithms (PBAs) are excellent search tools that allow searching space of parameters defined by problems under consideration. They are especially useful when it is difficult to define a differentiable evaluation criterion. This applies, for example, to problems that are a combination of continuous and discrete (combinatorial) problems. In such problems, it is often necessary to select a certain structure of the solution (e.g. a neural network or other systems with a structure usually selected by the trial and error method) and to determine the parameters of such structure. As PBAs have great application possibilities, the aim is to develop more and more effective search formulas used in them. An interesting approach is to use multiple populations and process them with separate PBAs (in a different way). In this paper, we propose a new multi-population-based algorithm with: (a) subpopulation evaluation and (b) replacement of the associated PBAs subpopulation formulas used for their processing. In the simulations, we used a set of typical CEC2013 benchmark functions. The obtained results confirm the validity of the proposed concept.

A New Approach to Detecting and Preventing Populations Stagnation Through Dynamic Changes in Multi-Population-Based Algorithms

“…In the [28], individual island training plans (both population algorithms and their parameters) are changed based on the optimization progress, which brought a significant improvement. In the paper [29], when stagnation was detected, various mechanisms were used to re-initialize those populations that performed the worst.…”