Optimizing Artificial Neural Network Learning Using Improved Reinforcement Learning in Artificial Bee Colony Algorithm

Abstract

Machine learning methods like artificial neural networks (ANNs) have many uses and applications. The loss function in artificial
neural networks (ANNs) is usually minimized using heuristic search optimization techniques. Nevertheless, the network’s
performance may be severely hindered if these strategies cause it to get trapped in local optima. enhanced R-ABC algorithm,
which stands for “improved reinforcement learning in the artificial bee colony,” is introduced as a new optimization strategy to
ANNs in this paper to address this difficulty. To increase the effectiveness of weight modification in ANNs and overcome the
constraints of heuristic search, the suggested technique is put forth. By including the properties of nearby food sources, this
novel method improves the discovery phase of the conventional R-ABC and increases the search capabilities for finding the
best solution. The enhanced R-ABC was tested against several artificial neural networks (ANNs) that use backpropagation with
stochastic gradient descent (SGD) and Adam optimizers, as well as other SI approaches including particle swarm optimization
(PSO) and conventional R-ABC. Across all benchmark datasets, the findings demonstrated that solutions were consistently
improved by both PSO and R-ABC. Unlike SGD and Adam, all SI methods routinely obtained F1-scores higher than 0.94 on
the iris dataset. Overall, the SI method performed better than the other optimization techniques for the other datasets. Applying
the enhanced R-ABC to ANNs yields better results than heuristic search optimization, particularly as the size of the network
increases, according to the findings. While SGD and Adam were able to achieve quicker execution times using TensorFlow, the
research indicates that PSO and enhanced R-ABC may enhance model correctness and efficiency. The optimization process and
ANNs’ capacity to find optimum solutions are both improved by modern SI methodologies. In complicated and high-dimensional
datasets, the performance and efficiency of artificial neural network (ANN) training may be greatly enhanced with the
use of improved R-ABC and PSO algorithms.