Development of the artificial hummingbird optimization algorithm and its application to feedforward neural networks

Many problems that were difficult or impossible to formulate mathematically have become solvable with the advancement of computers' ability to learn complex data patterns, establish relationships between events, and make decisions. The basis of this development lies artificial intelligence, artificial neural networks, and intelligent optimization algorithms inspired by nature. Metaheuristic methods aim to produce near-optimal solutions in complex search spaces by imitating the intuitive behavior of living organisms and are effectively used in many fields, from engineering to data mining. In this thesis, a new hybrid model, CLAHA (Chaotic-Initialized Light Hybrid AHA), is proposed, based on the Artificial Hummingbird Optimization Algorithm with various improvements, for training feedforward artificial neural networks. The improvements include a balanced distribution of the initial population using chaotic maps, preserving population diversity using mutation-based fine-tuning operators, and generating alternative individuals in opposing regions of the solution space using adversarial learning. Among the eight chaos maps tested, the Chebyshev map, which provided the best convergence performance, was selected for initial use. The proposed method was applied to optimize weight and bias parameters in the training of feedforward artificial neural networks and evaluated on 17 classification datasets. Experiments were conducted with 500 iterations, a population of 30 individuals, and 10 replicates; accuracy, F1-Score, AUC, MSE, and processing time metrics were used. The results showed that CLAHA achieved the highest accuracy and F1-Score in 13 of the 17 datasets, and the lowest error value in 10; in comparison with classical classifiers, it showed the best or equivalent performance in 12 of the 17 datasets. The overall optimization capability of the algorithm was also evaluated with 23 classical benchmark functions, CEC-2014 and CEC-2022 international test functions, and six engineering problems; CLAHA produced the best results in 15 out of 23 functions in CEC-2014, 15 out of 30 functions in CEC-2022, and 7 out of 12 functions in CEC-2022. The findings demonstrate that CLAHA significantly reduces the problem of getting stuck in local minima thanks to chaotic initialization and population diversity-preserving mechanisms and is an efficient and stable global optimization method for both classification and complex optimization problems.