Development of global real time numerical optimizer

• Main Author: Auditya Randi
• Other Authors: Suresh Sundaram
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/62669
• Institution: Nanyang Technological University

In this report, a new particle swarm optimization algorithm termed as Human Cognition Inspired Particle Swarm Optimization (HCIPSO) is proposed. HCIPSO adopts learning strategies inspired by human cognitive psychology and addresses the issue related to rotation variance properties of the search space. Studies in human cognitive psychology have indicated that the best planner regulates his strategies with respect to his current state and his perception of the best experiences from others. Rotation of the search space of a problem usually causes the dimensions of the problem to become non-separable where the optimum solution cannot be located by optimizing the objective function along each dimension separately. Based on these ideas, we propose four learning strategies for the HCIPSO algorithm namely self-regulating inertia weight, the self-perception on the global search direction, the rotation update strategy, and social direction guidance for the least performing particles. The self-regulating inertia weight is employed by the best particle for better exploration. Next, the group of the least performing particles is guided by group of elite particles categorized through the fitness values to achieve faster convergence rate. Lastly, the remaining particles will be randomly selected to employ the self-perception on the global search direction or the rotation update strategy to explore the rotation variance property of the search space. HCIPSO algorithm has been evaluated using the 28 benchmark functions from CEC2013. The results have been compared with seven state-of-the-art PSO variants such as Social Learning PSO (SL-PSO), Comprehensive Learning PSO (CLPSO), Self-Regulating PSO, etc. The proposed learning strategies in HCIPSO manage to achieve better convergence characteristics and provide better solutions in most number of the problems. Furthermore, a statistical analysis on performance evaluation of the different algorithms on CEC2013 benchmark functions indicates that HCIPSO is better than six out of seven selected algorithms with a 90% confidence level.