Feature: implement simulated campaigning for "hyper parameter tuning"

[matthewcarbone]

@ziatdinovmax as we discussed, I plan on implementing a simulated campaigning loop for tuning the "hyper parameters" of an optimization loop. I first want to learn this library inside and out so it might take some time. But anyways, the executive summary of the tasks at hand look something like this:

• Develop a Campaign class for storing the state of and running the campaign
• Parallelize the campaigning (since many simulations will have to be run); might want to consider mpi4py but more likely multiprocessing will be enough
• Implement a smart checkpoint-restart system for the same reason
• Write appropriate tests

[ziatdinovmax]

@matthewcarbone - sounds great, and I will be happy to help.

[matthewcarbone]

@ziatdinovmax Quick update: I have not forgotten about this. I have funding starting in October and I'll be building on this.

Btw, unrelated question (we can open a new issue if you want), but can gpax do batch sampling? I.e. instead of sequential experiments ("given the data at hand, find me the next experiment to maximize the acquisition function"), can gpax do "given the data at hand, find me the next q experiments that jointly maximize the acquisition function"?

[ziatdinovmax]

@matthewcarbone - Thanks for the update.
Yes, there is a batch-mode acquisition: https://github.com/ziatdinovmax/gpax/blob/main/gpax/acquisition/batch_acquisition.py

[ziatdinovmax]

On the "parallelize the campaigning": assuming this is a single program that runs with different input parameters, can this be done with JAX built-in tools for parallel evaluation?

[matthewcarbone]

It's funny I was thinking something similar, but I don't quite know how to do this. The tough part is that it's a combination of the continuous and bandit optimization. There's almost like a tree of decisions. For example, do you choose EI or UCB? If you choose UCB, you also need to choose beta. How does one go about optimizing over that space? I know it's possible, but I'm not sure how to implement it.

Btw, I also have concerns about speed. Fitting gpax to ~400 5-dimensional data points took quite a few minutes. I realize that's a lot of data, but I noticed that other codes are much faster. Is there any way we can speed up mcmc.run?

[ziatdinovmax]

One can use stochastic variational inference GP (viGP) or deep kernel learning (viDKL) for large datasets and high dimensions. The mcmc (or, more precisely, HMC with NUTS sampler) implementation is already dramatically faster than what pymc or pyro packages offer. I generally recommend it in situations where specific physics-based priors are available or one wants a detailed analysis of posterior distributions.

[matthewcarbone]

Is there a way to do this already in GPax?

Edit: whoops please disregard. 😁