• Comparative Study On Cooperative Particle Swarm Optimization Decomposition Methods for Large-scale Optimization

      Clark, Mitchell; Department of Computer Science
      The vast majority of real-world optimization problems can be put into the class of large-scale global optimization (LSOP). Over the past few years, an abundance of cooperative coevolutionary (CC) algorithms has been proposed to combat the challenges of LSOP’s. When CC algorithms attempt to address large scale problems, the effects of interconnected variables, known as variable dependencies, causes extreme performance degradation. Literature has extensively reviewed approaches to decomposing problems with variable dependencies connected during optimization, many times with a wide range of base optimizers used. In this thesis, we use the cooperative particle swarm optimization (CPSO) algorithm as the base optimizer and perform an extensive scalability study with a range of decomposition methods to determine ideal divide-and-conquer approaches when using a CPSO. Experimental results demonstrate that a variety of dynamic regrouping of variables, seen in the merging CPSO (MCPSO) and decomposition CPSO (DCPSO), as well varying total fitness evaluations per dimension, resulted in high-quality solutions when compared to six state-of-the-art decomposition approaches.
    • Efficient Merging and Decomposition Variants of Cooperative Particle Swarm Optimization for Large Scale Problems

      Douglas, Jay; Department of Computer Science
      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.
    • Landscape Aware Algorithm Configuration

      Dennis, Cody; Department of Computer Science
      The issue of parameter selection cannot be ignored if optimal performance is to be obtained from an algorithm on a specific problem or if a collection of algorithms are going to be compared in a fair manner. Unfortunately, adequately addressing the issue of parameter selection is time consuming and computationally expensive. Searching for appropriate control parameters generally requires much more time than actually solving the problem at hand due to the need to perform many complete runs of the target algorithm. The number of runs required to obtain thorough and equal coverage of the parameter space grows exponentially with the number of parameters. As a result, costs associated with parameter selection become a limiting factor in the scale of problems that can be investigated. The primary goal of this work is to reduce the costs of parameter selection. In pursuit of this goal, this thesis examines the use of neural networks to intelligently select appropriate control parameter values based on the characteristics of the problem at hand. Two general purpose approaches are evaluated: one that predicts a single set of control parameters to use throughout a run of the target algorithm; and, another that dynamically adjusts algorithm control parameters at run time. These approaches are examined in detail using the Particle Swarm Optimization algorithm. A comparison with state of the art automated tools for control parameter selection indicates that the cost of parameter selection can be significantly reduced.