GPU-accelerated Evolutionary Multiobjective Optimization Using Tensorized RVEA

TL;DR

TensorRVEA targets large-scale multiobjective optimization by leveraging GPU acceleration and full tensorization of data and operators. It transforms standard EMO components into tensor forms and implements a tensorized RVEA on EvoX, achieving up to thousands of times speedups. The approach is validated on large-scale numerical benchmarks (DTLZ) and applied to multiobjective neuroevolution in robotic control, showing superior speed and quality (HV, EU) compared to baselines. The work also demonstrates extensibility by swapping reproduction operators (GA, PSO, DE, CSO), indicating broad applicability to GPU-accelerated EMO workflows.

Abstract

Evolutionary multiobjective optimization has witnessed remarkable progress during the past decades. However, existing algorithms often encounter computational challenges in large-scale scenarios, primarily attributed to the absence of hardware acceleration. In response, we introduce a Tensorized Reference Vector Guided Evolutionary Algorithm (TensorRVEA) for harnessing the advancements of GPU acceleration. In TensorRVEA, the key data structures and operators are fully transformed into tensor forms for leveraging GPU-based parallel computing. In numerical benchmark tests involving large-scale populations and problem dimensions, TensorRVEA consistently demonstrates high computational performance, achieving up to over 1000$\times$ speedups. Then, we applied TensorRVEA to the domain of multiobjective neuroevolution for addressing complex challenges in robotic control tasks. Furthermore, we assessed TensorRVEA's extensibility by altering several tensorized reproduction operators. Experimental results demonstrate promising scalability and robustness of TensorRVEA. Source codes are available at \url{https://github.com/EMI-Group/tensorrvea}.

Figures (9)

• Figure 1: Illustration of reference vector guided selection in RVEA. The population is partitioned into subpopulations by associating each with a reference vector, according to the angle-penalized distance (APD) metric.
• Figure 2: Example of objective space partitioning and APD calculation for $n=6$, $r=3$. Left: objective space partitioning. Right: APD calculation.
• Figure 3: Performance comparison on CPU and GPU platforms when scaling the population size on DTLZ1 problem. Results highlight the significant speedup achieved by using GPU over CPU, with TensorRVEA on GPU showing a remarkable 1528.0$\times$ speedup at the largest population size examined.
• Figure 4: Performance comparison on CPU and GPU platforms when scaling the problem dimension on DTLZ1 problem. Results highlight the significant speedup achieved by using GPU over CPU, with TensorRVEA on GPU showing a remarkable 1042.0$\times$ speedup at the largest problem dimension examined.
• Figure 5: Comparison of mean IGD values and 95% confidence intervals for TensorRVEA and RVEA on DTLZ1-DTLZ4 problems.