Avaliação de Desempenho de Metaheurísticas Usando o Paralelismo do TensorFlow e PyTorch

Abstract

Computational optimization has become a critical challenge in engineering and artificial intelligence, especially in parallel computing, due to the high complexity of solutions and the intensive demand for computational resources. Metaheuristics, such as Particle Swarm Optimization (PSO), Cuckoo Search (CS), Differential Evolution (DE), and Genetic Algorithms (GA), are effective approaches that combine local and global search to explore large solution spaces. However, executing these techniques can be computationally expensive, requiring efficient parallelization strategies. This study evaluates the performance of parallelized metaheuristics using TensorFlow and PyTorch on benchmark optimization functions, analyzing metrics such as speedup and efficiency between the CPU and GPU. Techniques such as tf.distribute.MirroredStrategy, which replicates tensors across multiple devices within the strategy.scope, and, in PyTorch, the device = torch.device(’cuda’) strategy, allowing operations like torch.rand, torch.cat, torch.argsort, and torch.where to be efficiently distributed, were employed. The experiments demonstrate the impact of parallelism on the search for global and local solutions, highlighting the role of GPUs and parallel architectures in computational optimization. This technical analysis provides insight into the selection of the most suitable technologies for critical applications where parallelism is justified by the need to solve complex and high-dimensional problems