GPU-accelerated Evolutionary Multiobjective Optimization Using Tensorized RVEA

Zhenyu Liang, Tao Jiang, Kebin Sun, Ran Cheng

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.

Introduction

Many real-world optimization problems are multiobjective, requiring a balance between conflicting goals—such as a robot being both fast and energy-efficient. Evolutionary Multiobjective Optimization (EMO) algorithms excel at finding a "Pareto Front" of optimal trade-offs. However, as problems grow in scale (Large-Scale Multiobjective Optimization, or LSMOPs), traditional EMO algorithms hit a computational wall. Most are designed for CPUs, which process tasks sequentially, while GPUs—the gold standard for parallel processing—have been difficult to utilize because EMO logic (sorting, partitioning, and selection) does not naturally fit matrix-math architectures.

Methodology: The Power of Tensorization

TensorRVEA addresses this by being a fully "tensorized" version of the Reference Vector Guided Evolutionary Algorithm, built on the EvoX framework and powered by JAX. The core innovation is the transformation of traditional evolutionary operators into high-dimensional tensor operations:

  1. Tensorized Data Structures: The entire population, objective scores, and reference vectors are represented as multi-dimensional tensors rather than lists of individuals.
  2. Tensorized Operators: Core steps like Simulated Binary Crossover (SBX), Polynomial Mutation, and selection are rewritten into pure tensor math.
  3. Parallel Selection: RVEA partitions the objective space into independent subspaces. TensorRVEA handles these partitions simultaneously across thousands of GPU cores using JAX’s vmap (vectorized map) function, making the selection logic "embarrassingly parallel."
The tensorized workflow for objective space partitioning and Angle-Penalized Distance (APD) calculation.

Experimental Evaluation

The researchers evaluated TensorRVEA using two primary benchmarks:

  • Numerical Benchmarks (DTLZ1-DTLZ4): Standard synthetic problems used to test convergence and diversity.
  • Multiobjective Neuroevolution (Brax): High-performance robotic control tasks (e.g., MoHalfCheetah, MoHopper) where the algorithm evolves neural network weights to balance rewards against control costs.

Key Results and Insights

TensorRVEA achieved staggering performance gains. Compared to standard CPU implementations, it reached a 1528× speedup for large populations (16,384 individuals) and a 1042× speedup for high-dimensional decision spaces (over 260,000 variables).

Computational performance of TensorRVEA vs. standard RVEA across varying population sizes, highlighting the 1528x speedup.

Beyond raw speed, the algorithm is significantly more efficient in time-constrained scenarios. In a fixed window of just 90 milliseconds, TensorRVEA achieved significantly better convergence than CPU versions, which could barely complete a few generations. In robotic tasks, it consistently outperformed classic algorithms like NSGA-II, finding a wider and more accurate range of behaviors.

Conclusion and Future Work

While TensorRVEA is designed for large-scale problems—and may face overhead on very small populations—it demonstrates that tensors are a powerful abstraction for population-based search. By using JAX for just-in-time (JIT) compilation, the authors have paved the way for a new generation of GPU-native evolutionary frameworks that can handle the complexity of modern large-scale optimization and neuroevolution.

Summary updated

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}.
BibTeX
@misc{liang2024gpu,
  author = {Zhenyu Liang and Tao Jiang and Kebin Sun and Ran Cheng},
  title = {{GPU-accelerated Evolutionary Multiobjective Optimization Using Tensorized RVEA}},
  year = {2024},
  eprint = {2404.01159},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.01159},
}

Loading related content...

Figure