Efficient Merging and Decomposition Variants of Cooperative Particle Swarm Optimization for Large Scale Problems

Abstract

For large-scale optimization problems (LSOPs), an increased problem size reduces performance by both increasing the landscape complexity, as well as exponentially increasing the search space size. These contributing factors make up the "curse of dimensionality", which is addressed either by improving the search operator of the meta-heuristic or decomposing the high-dimensional problem into smaller sub-problems. Unfortunately, non-separable LSOPs contain a scaling number of variable dependencies which should be optimized together but are often separated into different sub-problems due to insufficient grouping strategies. Various particle swarm optimization (PSO) techniques have been proposed in order to address these LSOPs, either through the improvement of search operators or utilizing decomposition. However, there is a lack of comparison between them showing which PSO variant performs best for specific types of LSOPs. Additionally, decomposition variants which utilize a cooperative PSO (CPSO) approach still struggle to properly group related variables in more difficult non-separable multimodal problems. In an attempt to better optimize these non-separable LSOPs, this thesis introduces two new adaptive decomposition and merging CPSO algorithms, referred to as DCPSO2 and MCPSO2 respectively, which offer a new regrouping strategy by adaptively splitting and merging stagnated sub-swarms according to the their fitness. These algorithms proposed in this thesis are then compared against existing CPSO variants in order to establish the best decomposition-based PSO algorithm for LSOPs. Results show that the decomposition and merging variants are able to perform competitively with previously well-established CPSO algorithms for large-scale problems across all problem classes. DCPSO ranks the highest in terms of accuracy across all non-separable problems while MCPSO and MCPSO2 prove to have the fastest convergence amongst all algorithms.