Hey there, fellow AI enthusiasts! Today, I want to share an exciting development in the world of deep reinforcement learning. As an independent researcher, I’ve been exploring ways to optimize training in noisy environments, and I’m thrilled to introduce Ano, a new optimizer that tackles this challenge head-on.
The idea behind Ano is simple yet powerful: decouple the magnitude of the gradient from the direction of the momentum. This approach aims to make training more stable and faster in environments that are common in deep RL settings.
I’ve made my preprint and source code available for feedback and review. I’d love to hear your thoughts on the method and the clarity of the writing. Your input will help me refine Ano and make it more effective.
You can check out the preprint and source code here: https://zenodo.org/records/16422081. If you’re interested in trying Ano out, you can install it via pip: pip install ano-optimizer. I’ve also set up a GitHub repository for experiments: https://github.com/Adrienkgz/ano-experiments.
As this is my first research contribution, I’m eager to hear your feedback, suggestions, or constructive criticism. Your input will be invaluable in helping me improve Ano.
Additionally, I’m seeking endorsement to publish my preprint on arXiv. If you’re familiar with the platform and feel comfortable endorsing my work after reviewing the paper, I’d greatly appreciate it.
Thanks for taking the time to read about Ano, and I look forward to your feedback!