Evotorch

EvoTorch is an open source evolutionary computation library built directly on PyTorch and developed at NNAISENSE. It targets optimization problems that gradient descent cannot handle, including black box optimization over continuous or discrete spaces, reinforcement learning tasks and supervised learning tasks, all expressed through one problem and solver interface. It is written for researchers and machine learning engineers who need population based search rather than backpropagation. The library ships distribution based algorithms covering policy gradients with parameter based exploration, exponential and separable natural evolution strategies, covariance matrix adaptation evolution strategies and the cross entropy method, alongside population based methods including a genetic algorithm that also handles multiple objectives, cooperative synapse neuroevolution, and a multi dimensional archive of phenotypic elites. Because every algorithm is implemented in PyTorch, populations are vectorized and fitness evaluation runs on CUDA hardware when a GPU is available. Ray support spreads a single search across multiple cores and multiple machines in a cluster. Logging plugs into common experiment tracking tools, and the modular design is intended to let researchers drop in their own operators and algorithms. There is no pricing. The library is free, released under the Apache License 2.0, and installed with a single pip command. Documentation, tutorials and example notebooks are published on a separate documentation site alongside the public repository. One practical note: the marketing site is a thin landing page with little content, so the documentation site and the public repository are where the real reference material lives. Copyright notices on the documentation credit NNAISENSE, the Swiss AI company behind the project.