Towards Efficient Evaluation of Evolutionary Transfer Optimization: Case Studies on Task-Parameterized Applications

Summary

The paper addresses runtime growth from problem evaluation in evolutionary transfer optimization (ETO), where repeated evaluation across many tasks or high-resolution evaluations dominates wall-clock time. It organizes evaluation scaling into two regimes: cross-task scaling (number of tasks K) and within-task scaling (per-evaluation workload η). Two case studies show that application-specific serial evaluations can be rewritten into mathematically equivalent parallel matrix forms. First, kinematic-arm multi-task optimization: the joint-by-joint matrix recursion is replaced by a prefix-sum accumulation matrix of cumulative link angles plus direct trigonometric projections, yielding a 256.72× end-to-end speedup at K=5000 tasks (100 generations, N=16, D=60) with mean absolute fitness discrepancy 2.561×10⁻⁸. Second, B-spline trajectory sequential transfer optimization: pointwise spline evaluation is replaced by a dense R×C blending matrix (Q=BP), a difference matrix for path length, and a joint segment-obstacle intersection matrix, yielding a 93.91× end-to-end speedup at R=6000 samples with discrepancy 1.101×10⁻⁵. Experiments use EvoX on a single RTX 3090; code and scripts are open source. Limitations: only two continuous applications; reformulations are application-specific; speedups are measured against serial reference implementations whose GPU status is unstated; optimization quality is thinly characterized (MS-STO reached collision-free trajectories in only 4/10 runs); memory overhead is acknowledged but not measured.