Reproduce “EvoGrad: Metaheuristics in a Differentiable Wonderland” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Beatrice F.R. Citterio, Andrea Tangherloni
EvoGrad is a framework that makes population-based metaheuristics fully differentiable, enabling end-to-end gradient optimization of both solutions and algorithmic hyperparameters. This approach allows differentiable versions of CMA-ES and PSO to match or exceed the performance of their classical counterparts on continuous optimization benchmarks and neural network training, while benefiting from GPU acceleration.
Key insights
- Transforming stochastic, discrete operations in metaheuristics into continuous, differentiable counterparts is possible using techniques like the reparameterization trick and Gumbel-Softmax/Sigmoid relaxations.
- Making all algorithmic hyperparameters learnable via backpropagation enables automatic, adaptive tuning, removing the need for manual configuration.
- A unified gradient and exploration loop can interleave stochastic population-based search with gradient-driven parameter adaptation, executed efficiently on GPUs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.