[disclosure: I work in this space and make another kind of optimizer.]
Sorry to talk about a competing product here...
We've been applying deep reinforcement learning to various scheduling problems (which are hard!), and it has shown performance gains of between 10% and 32% on various use cases.
The thing that RL can do, which mathematical solves sometimes struggle with, is generalize for highly variable data, and be updated quickly. (You can retrain the policy without rewriting anything manually, unlike MIPS.)
Sorry to talk about a competing product here...
We've been applying deep reinforcement learning to various scheduling problems (which are hard!), and it has shown performance gains of between 10% and 32% on various use cases.
The thing that RL can do, which mathematical solves sometimes struggle with, is generalize for highly variable data, and be updated quickly. (You can retrain the policy without rewriting anything manually, unlike MIPS.)
We're doing that here:
https://pathmind.com/
We're automating several steps in RL, including choosing the architecture and the hyperparameters.
https://pathmind.com/why-pathmind/
Anyone who would like to learn more should feel free to message me.