GPU-Accelerated GP-GOMEA Cracks Feynman Equations in 4 Hours
Achieving orders of magnitude more evaluations via GPU parallelism for symbolic regression.
GP-GOMEA is a leading evolutionary algorithm for symbolic regression, prized for producing small, interpretable models. However, its high computational cost has limited its use on large datasets and complex expressions. Meanwhile, deep learning has thrived due to GPU parallelism—a resource evolutionary algorithms have largely underutilized. In the paper "GP-GOMEA with GPU-Based Fitness Evaluations: Design and Performance Analysis," Jasper Post and his team take the first major step toward fully GPU-accelerated GP-GOMEA. They introduce a novel GPU-friendly representation of GP-GOMEA's template-based individuals and a parallel evaluation strategy that exploits population-based search's inherent parallelism.
This approach dramatically increases evaluation throughput, enabling orders of magnitude more evaluations within the same time budget. On four standard symbolic regression benchmarks, the increased capacity yields significant performance improvements, especially for larger datasets and population sizes. Furthermore, the ability to efficiently evaluate much larger datasets and more complex templates allows systematic analysis of expression difficulty. For the first time, this enables a problem-agnostic evolutionary algorithm to reliably regress one of the largest Feynman equations within just four hours—a task previously infeasible.
- GPU acceleration provides orders of magnitude more fitness evaluations per time unit compared to CPU-based GP-GOMEA.
- Performance gains are most pronounced on larger datasets and larger population sizes in symbolic regression benchmarks.
- For the first time, a problem-agnostic evolutionary algorithm regressed a large Feynman equation within four hours using GPU parallelism.
Why It Matters
Brings evolutionary symbolic regression up to GPU speed, enabling interpretable model discovery on previously intractable problems.