Scholarlink

In this study, we investigate the efficiency of optimizer structures based on modifications of the ant colony optimization (ACO) method applied to parametric problems, with special attention to the time required to find the values of the objective function. The use of separate model threads implies parallel, simultaneous execution of optimization and computation of the objective function values. Modifications of ACO for solving parametric problems are robust to parametric graph state shifts. We study modifications of ACO for parametric optimization running on GPUs, which allow separating the computation of model threads and the optimizer thread within personal computers, as well as applying ACO modifications to supercomputer optimization. Execution of a matrix-optimized modification of ACO on GPUs for parametric optimization on an NVIDIA Tesla V100 16GB SXM2 computer requires about 1.5 ms per parameter/layer of the parametric graph, including for very large-scale problems. For heterogeneous execution of the optimization flow, matrix normalization, addition, and evaporation of pheromone weights are performed on the CPU using the SSE/AVX instructions, and the search for paths by agent ants is performed on the GPU. With a significant transmission time, a pipeline processing mechanism is proposed, which is effective when the search times for criteria values, transmission, and calculation of the objective function values are equal. Various topologies of the distributed computer are investigated.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimization of the Structure of the Solution of Parametric Problems by Methods of Modification of Ant Colonies Taking into Account Time Factors

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yurii Titov

摘要

In this study, we investigate the efficiency of optimizer structures based on modifications of the ant colony optimization (ACO) method applied to parametric problems, with special attention to the time required to find the values of the objective function. The use of separate model threads implies parallel, simultaneous execution of optimization and computation of the objective function values. Modifications of ACO for solving parametric problems are robust to parametric graph state shifts. We study modifications of ACO for parametric optimization running on GPUs, which allow separating the computation of model threads and the optimizer thread within personal computers, as well as applying ACO modifications to supercomputer optimization. Execution of a matrix-optimized modification of ACO on GPUs for parametric optimization on an NVIDIA Tesla V100 16GB SXM2 computer requires about 1.5 ms per parameter/layer of the parametric graph, including for very large-scale problems. For heterogeneous execution of the optimization flow, matrix normalization, addition, and evaporation of pheromone weights are performed on the CPU using the SSE/AVX instructions, and the search for paths by agent ants is performed on the GPU. With a significant transmission time, a pipeline processing mechanism is proposed, which is effective when the search times for criteria values, transmission, and calculation of the objective function values are equal. Various topologies of the distributed computer are investigated.