Design of an Adaptive Push-Repel Operator for Enhancing Convergence in Genetic Algorithms

Genetic Algorithms (GAs) are demonstrated to be successful in solving problems pertaining to the field of engineering, physics, medicine, finance and many more. The efficacy of GAs lies in its efficiency at exploring complex design-space with black-box constraints and reach the optimal regions defined by functions of unknown fitness landscapes (or in other words, black-box optimization functions). Depending on the nature of the problem, the design-space can have continuous, discrete or mixed (continuous and discrete) set of design-variables. The exploration in this design-space is conducted through a population of individuals and is primarily driven by three operations –selection, recombination (or crossover) and mutation. The exploitation aspect of a GA search is obtained by its selection operation, while crossover and mutation operations deal with the exploration aspect for generating new solutions in the search space. In this study, an attempt has been made to balance the two aspects by designing a generic push operator which introduces an extra level of exploitation in the algorithm by biasing the creation of solutions near the best-so-far solution. In addition to standard search operators, an additional diversity maintaining repel operator is introduced to balance the exploitation-exploration issue. Simulations are performed to understand the effect of an adaptive push-repel GA on different fitness landscapes for both unconstrained and constrained optimization problems. The results are promising and encourage their extensions to other evolutionary algorithms.