A (mostly faithful) implementation of the 2016 paper Learning to Optimize. You can find a more casual explanation in this blog post.
The main idea behind the paper is simple: instead of using handcrafted optimization algorithms, why not learn them? The paper presents a framework for learning optimization algorithms for machine learning objectives. They cast this as a reinforcement learning problem, where the algorithm is the policy, and the reward is the negative objective function. Check out the blog post or paper for more details.
Disclaimer: This was a final project for a class. You can find the hastily written report here, which contains details about implementation and a few performance plots. Unfortunately, there wasn't time to implement the Guided Policy Search algorithm used in the original work (more here and here), so we tested with standard RL algorithms like A2C and PPO instead. We obtained similar performance as the original paper on most objectives, but performed worse when optimizing the 2-layer MLP. However we also performed little hyperparameter tuning, and think performance could be significantly improved.
I hope this can be a useful reference for others looking to explore optimizer learning. And of course, feel free to reach out to me!
To install requirements, run:
$ pip install -r requirements.txtTo run the code, check out the notebook experiments.ipynb.
Format with make format, lint with make lint.