EvoGrad: Metaheuristics in a Differentiable Wonderland

TL;DR

EvoGrad introduces a differentiable framework combining evolutionary algorithms and swarm intelligence with gradient-based methods, leading to improved optimization performance in complex search spaces and neural network training.

Contribution

It presents the first unified differentiable framework for EC and SI, enabling end-to-end gradient optimization of traditionally non-differentiable operators.

Findings

Differentiable EC and SI outperform traditional algorithms in benchmark tests.

EvoGrad enhances neural network training efficiency.

The framework sets a new standard for hybrid optimization methods.

Abstract

Differentiable programming has revolutionised optimisation by enabling efficient gradient-based training of complex models, such as Deep Neural Networks (NNs) with billions and trillions of parameters. However, traditional Evolutionary Computation (EC) and Swarm Intelligence (SI) algorithms, widely successful in discrete or complex search spaces, typically do not leverage local gradient information, limiting their optimisation efficiency. In this paper, we introduce EvoGrad, a unified differentiable framework that integrates EC and SI with gradient-based optimisation through backpropagation. EvoGrad converts conventional evolutionary and swarm operators (e.g., selection, mutation, crossover, and particle updates) into differentiable operators, facilitating end-to-end gradient optimisation. Extensive experiments on benchmark optimisation functions and training of small NN regressors reveal that our differentiable versions of EC and SI metaheuristics consistently outperform traditional, gradient-agnostic algorithms in most scenarios. Our results show the substantial benefits of fully differentiable evolutionary and swarm optimisation, setting a new standard for hybrid optimisation frameworks.