evotorch

Traditional evolutionary algorithms (genetic algorithms, evolution strategies, CMA-ES) are implemented in NumPy with CPU-bound execution; EvoTorch reimplements these algorithms using PyTorch tensors, enabling them to run on GPU hardware at scale and to evolve neural network weights directly in GPU memory alongside PyTorch's autograd infrastructure.

The library supports a range of evolutionary algorithmsPGPE, CMA-ES, SNES, XNES, and custom variantsalongside multi-objective optimization methods like NSGA-II for problems with competing objectives.

EvoTorch's problem definition API supports both black-box optimization (where gradient information is unavailable) and hybrid approaches that combine gradient-based and gradient-free optimization.

It integrates with Ray for distributed evolution across multiple machines, enabling population sizes and evaluation throughputs that single-machine implementations cannot achieve.

Reinforcement learning researchers using neuroevolution as an alternative to gradient-based policy optimization, engineers optimizing non-differentiable objectives (game playing, robotics control, combinatorial problems), and practitioners exploring neural architecture search with evolutionary methods use EvoTorch to access GPU-accelerated evolutionary computation without low-level CUDA programming.

The PyTorch foundation means evolved solutions integrate naturally with the broader deep learning ecosystemevolved network weights can be fine-tuned with backpropagation, and torch.nn modules serve directly as individuals in the evolutionary population.