Deterministic population initialization for multi-objective mapping of multiple applications in network-on-chip-based many-core systems

This article investigates multi-objective optimization for multi-application mapping in Network-on-Chip (NoC)-based many-core systems. Although evolutionary algorithms have been widely used for NoC task mapping, the combined use of deterministic population initialization and multi-objective optimization in multi-application scenarios remains insufficiently explored. The problem formulation considers three objectives: communication cost, used as a proxy for communication energy, Load Balance, and Fault Tolerance, evaluated through two bi-objective formulations: Fault Tolerance × Energy and Load Balance × Energy. We developed a modular high-level framework integrating synthetic and benchmark application generation, multi-objective optimization via PlatEMO, and result visualization. The framework supports 6 × 6, 8 × 8, and 10 × 10 NoC grids and different multi-application scenarios. Four evolutionary algorithms—NSGA-II, NSGA-III, PESA-II, and SPEA2—were evaluated with random initialization and four deterministic initialization strategies: Horizontal Raster, Horizontal Snake, Diagonal Raster, and Diagonal Snake. The experiments considered up to seven concurrent applications and 50 independent runs per scenario. The NSGA-III family achieved the lowest representative distance-to-origin value in 22 of 24 scenarios (approximately 92%), while the deterministic NSGA-III variants achieved the lowest value in 15 scenarios (62.5%). Friedman and Iman–Davenport tests followed by Holm's post-hoc procedure indicated that NSGA-III obtained significantly better ranks than NSGA-II, SPEA2, and PESA-II. However, no statistically significant differences were found between the deterministic NSGA-III variants and the original NSGA-III baseline. These findings indicate that deterministic initialization can improve representative solutions in some scenarios, but its benefits are scenario-dependent.