Time Varying vs. Fixed Acceleration Coefficient PSO Driven Exploration during High Level Synthesis: Performance and Quality Assessment

Abstract

The performance of particle swarm optimization (PSO) greatly depends upon the effective selection of vital tuning metric known as acceleration coefficients (especially when applied to design space exploration (DSE) problem) which incorporates ability to clinically balance between exploration and exploitation during searching. The major contributions of the paper are as follows: a) A novel analysis of two variants of acceleration coefficient (hierarchical time varying acceleration coefficient vs. Constant acceleration coefficient) in PSO and their impact on convergence time and exploration time in context of multi objective (MO) DSE in HLS. The analysis assists the designer in pre-tuning the acceleration coefficient to an optimal value for achieving better convergence and exploration time before DSE initiation, b) A novel performance comparison of PSO driven DSE (PSO-DSE) with previous works based on quality metrics for MO evolutionary algorithms such as generational distance, maximum pareto-optimal front error, spacing, spreading and weighted metric. When two variants of acceleration coefficients (constant and time varying) were compared, it was revealed from the results that the PSO-DSE has on average 9.5% better exploration speed with constant acceleration coefficient as compared to hierarchical time varying acceleration coefficient. Further, with setting of constant acceleration coefficient, the PSO-DSE produces results with efficient generational distance, maximum pareto-optimal front error, spacing, spreading and weighted metric as compared to previous approaches.