An extension of the ACOR algorithm with time-decaying search width, with application to neural network training
Ashraf M. Abdelbar, Khalid M. Salama · 2016
ACOℝis a fairly-recent Ant Colony Optimization (ACO) algorithm for continuous problem domains. ACOR's search width parameter £ controls the extent to which the search is concentrated around the best solutions encountered so far. In this paper, we propose a variation of ACOℝ, in which ACOℝ's search width parameter decays over time. This is analogous to the common strategy, in Particle Swarm Optimization, of decreasing the inertia parameter over time. We evaluate our proposal in the context of neural network training using 36 popular datasets. We find that our approach produces solutions that are better, to a statistically significant extent, than the standard ACOℝalgorithm.