Evolutionary artificial bee colony for neural networks training
Mário Tasso Ribeiro Serra Neto, Marco Antônio Florenzano Mollinetti, Rodrigo Lisbôa Pereira · 2017
The Artificial Bee Colony is a popular simple and efficient bee inspired metaheuristic that showed good performance on real valued optimization problems. To improve its local search capabilities, a modified version of it, called ABC+ES was proposed. The new algorithm employs reproduction and mutation operators found in Evolutionary Strategies as well as a reference to the global best from the Particle Swarm Optimization, and has obtained better results in the same field of problems than the Bee Colony. However, testing was only made with problems that have a relatively small number of decision variables. With the intention of exploring the deep learning field, that features domains of high dimensionality, the following paper introduces an adaptation of the Artificial Bee Colony + Evolutionary Strategies for training neural networks. Performance is measured on six representative benchmarks that range from hundreds to thousands of connection weights against other well-known techniques such as exact methods and metaheuristics, as well as the original artificial bee colony. Results show that the approach performed on par with the artificial bee colony for minor scale problems, while it had a better outcome than every other technique on benchmarks with over 1,000 connection weights and multiple outputs. Further study is necessary to confirm whether the approach can be a viable alternative for training networks with very high dimensionality and various outputs aside from gradient methods.