Synergizing fitness learning with proximity-based food source selection in artificial bee colony algorithm for numerical optimization

Abstract

Evolutionary computation (EC) paradigm has undergone extensions in the recent years diverging from the natural process of genetic evolution to the simulation of natural life processes exhibited by the living organisms. Bee colonies exemplify a high level of intrinsic interdependence and co-ordination among its members, and algorithms inspired from the bee colonies have gained recent prominence in the field of swarm based metaheuristics. The artificial bee colony (ABC) algorithm was recently developed, by simulating the minimalistic foraging model of honeybees in search of food sources, for solving real-parameter, non-convex, and non-smooth optimization problems. The single parameter perturbation in classical ABC resulted in fairly commendable performance for simple problems without epistasis of variables (separable). However, it suffered from narrow search zone and slow convergence which eventually led to poor exploitation tendency. Even with the increase in dimensionality, a significant deterioration was observed in the ability of ABC to locate the optimum in a huge search volume. Some of the probable shortcomings in the basic ABC approach, as observed, are the single parameter perturbation instead of a multiple one, ignoring the fitness to reward ratio while selecting food sites, and most importantly the absence of environmental factors in the algorithm design. Research has shown that spatial environmental factors play a crucial role in insect locomotion and foragers seem to learn the direction to be undertaken based on the relative analysis of its proximal surroundings. Most importantly, the mapping of the forager locomotion from three dimensional search spaces to a multidimensional solution space calls forth the implementation of multiple modification schemes. Based on the fundamental observation pertaining to the dynamics of ABC, this article proposes an improved variant of ABC aimed at improving the optimizing ability of the algorithm over an extended set of problems. The hybridization of the proposed fitness learning mechanism with a weighted selection scheme and proximity based stimuli helps to achieve a fine blending of explorative and exploitative behaviour by enhancing both local and global searching ability of the algorithm. This enhances the ability of the swarm agents to detect optimal regions in the unexplored fitness basins. With respect to its immediate surroundings, a proximity based component is added to the normal positional modification of the onlookers and is enacted through an improved probability selection scheme that takes the T/E (total reward to distance) ratio metric into account. The biologically-motivated, hybridized variant of ABC achieves a statistically superior performance on majority of the tested benchmark instances, as compared to some of the most prominent state-of-the-art algorithms, as is demonstrated through a detailed experimental evaluation and verified statistically.