This paper addresses the evaluation bottleneck in Evolutionary Transfer Optimization (ETO) when dealing with large collections of related tasks. As the number of tasks grows, problem‑side evaluation often dominates runtime. We decompose evaluation scaling into two levels:
- Number of evaluated tasks – each task requires an independent evaluation in multi‑task optimization.
- Workload within a single task – the internal computation of each task also grows with problem size.
For both levels, we propose parallel‑friendly reformulations in two task‑parameterized applications.
-
Matrix‑recursive kinematic‑arm evaluation: The original recursive traversal of the kinematic chain is hard to parallelize. We rewrite it using an accumulation‑matrix that stores cumulative link directions, allowing all joint poses to be computed via a single batch of matrix multiplications, enabling massive parallelism.
-
Pointwise B‑spline trajectory evaluation (sequential transfer optimization): The conventional approach computes spline values and collision checks point‑by‑point, leading to linear growth with trajectory points. By introducing a blending‑matrix representation, trajectory generation and collision detection are merged into matrix multiplications, also suitable for GPU or multi‑core execution.
Both reformulations preserve numerical fidelity while delivering end‑to‑end speedups of $256.72\times$ and $93.91\times$, respectively. The results demonstrate that problem‑side reformulation is a practical route to scalable ETO. All implementation code and experimental scripts are released as open source to facilitate reproducibility and reuse.
Review: By converting inherently serial evaluation steps into matrix‑based linear algebra operations, the work provides a clear pathway to accelerate ETO on large‑scale task sets. Open‑sourcing the code further enhances reproducibility and community impact.