A hardware-accelerated Python library for running Quality-Diversity and neuroevolution algorithms in minutes instead of days.
QDax is a Python library that accelerates Quality-Diversity and neuroevolution algorithms using hardware accelerators like GPUs and TPUs. It solves the problem of extremely long runtimes for these population-based optimization methods, enabling researchers to run experiments in minutes instead of days or weeks. The framework is built on JAX for parallelization and is designed to be flexible and extensible for various research problems.
Researchers and practitioners in machine learning, robotics, and evolutionary computation who work with Quality-Diversity optimization, neuroevolution, or need to run large-scale population-based algorithms efficiently.
Developers choose QDax for its unparalleled speedup of QD algorithms through hardware acceleration, its comprehensive implementation of state-of-the-art algorithms, and its design as an extensible research framework that facilitates rapid prototyping and experimentation.
Accelerated Quality-Diversity
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Leverages JAX for automatic differentiation and vectorization to run on GPUs/TPUs, transforming algorithms that took weeks on CPUs into minute-long experiments.
Implements a wide range of core QD algorithms like MAP-Elites and CMA-ME, plus baselines such as NSGA2 and PBT for robust research comparisons.
Built with modularity and extensibility in mind, allowing easy adaptation to new problems and algorithms, as emphasized in its research-focused philosophy.
Provides ready-to-use QD tasks like robotic arm control, enabling immediate experimentation without the need to build environments from scratch.
Requires familiarity with JAX's functional programming and automatic differentiation, which can be a significant barrier for researchers accustomed to imperative frameworks like PyTorch.
Optimal performance is tightly coupled to GPU/TPU availability; on CPU-only systems, the touted speed advantages are drastically reduced.
Primarily designed for experimentation, lacking built-in tools for model deployment or seamless integration into production machine learning pipelines.