evotorch

Why we included this project

When backpropagation is not an option, because reward signals are discontinuous, search spaces are discrete, or gradients are simply impractical to compute, EvoTorch offers a practical alternative. The library packages evolutionary algorithms like PGPE, XNES, and CMA-ES as building blocks that operate directly on PyTorch tensors, so you can evolve network weights and policies without writing your own optimizer plumbing. Everything is tensor-based and parallelized, scaling across CPUs, GPUs, and multiple nodes, which suits reinforcement learning and neuroevolution workloads well. Researchers experimenting with novel learning rules and engineers tackling black-box optimization both get one coherent API with solid documentation, and the PyTorch foundation means it drops into existing model and training code with little friction.