Distributed Swarm Attention Learning Optimizer With Sensitive Search Transition for Large-Scale Optimization

Owing to high dimensionality and hence high computational cost, large-scale optimization problems (LSOPs) bring tremendous challenges to large-scale optimization algorithms (LSOAs). Faced with LSOPs, the sensitive search transition between exploration and exploitation is crucial, as it can effectively prevent redundant search. However, existing LSOAs are plagued by their inability to assess the evolution status with low computational cost. Besides, as the dimensionality of optimization problems increases, the search space expands exponentially. It is imperative to expedite search within such a vast search space. To this end, herein a distributed swarm attention learning optimizer with sensitive search transition (DSALO-SST) is proposed. DSALO-SST leverages a master-slave model, where the master and each slave build archives to store the historical optimal individuals of the population and related subpopulation, respectively. In DSALO-SST, we design an aggregation status indicator (ASI) which focuses on an archive instead of the correlated slave, reducing the computational cost. Based on ASI, a steady state indicator (SSI) is defined by incorporating historical statistical information. Through assessing the evolution status of an archive by ASI and SSI, the sensitive search transition can be conducted. Additionally, swarm attention learning (SAL) is devised to generate a composite learning target for each individual to be updated via allocating different attentions to the selected superior objects. Guided by the composite learning target that integrates multiple superior objects, the individuals' search can be expedited. The experimental results, particularly the comprehensive comparisons with seven state-of-the-art LSOAs, validate that DSALO-SST significantly improves the performance in handling LSOPs.