Simple BO Loop stuck on maximum bound.

[Friedemannn]

hi there,

I'm trying to optimize a fbpic simulation in a simple 1D BO loop.
But for some reason, I only get the maximum of my bounds as candidates from my UpperConfidenceBound.
At least for one of the x variables that I'm playing around with.
I currently have two versions, in one the x variable is nitrogen density (bounds:5e+22,1.5e+23) in the other it's laser focus (bounds:0.0001,0.0015)
In the laser focus one, I only get 0.0015 as a candidate and no exploration. In the other one it seems to work fine.

I already played around with the UCB beta, trying 0.1, 0.5 and 0.9. I also changed num_restarts (10 and 80) and raw_samples (256 and 4096). This didn't change anything.

Here are the two notebooks, there are also some visualizations in there, so you can see, that the maximum is not at 0.0015. Sorry for the very messy code.

bay_opt-simple-laser_foc-GH.ipynb
bay_opt-simple-GH.ipynb

Do you have some idea, what I could do to try to fix this?

[Balandat]

Are you trying to minimize or maximize your objective here? Is there something weird going on with the signs of max_chrg_in_bounds? These are positive at the beginning of the notebook but the values output later have a negative sign:

Somewhat unrelated to your questions: For the bay_opt-simple-GH.ipynb example, what is going on with the model fit? Between 0.6 and 0.8 you have these massive spikes that seem very wrong from eyeballing the observed data.

[Friedemannn]

Thanks for the reply!
I'm trying to maximize the electron charge, which is negative. That's why I'm using only absolute values for the tensor max_chrg_in_boundsT. For the array max_chrg_in_bounds (without T) I'm using the real simulations results. As far as i can see, this array is never used in the BO loop, so it shouldn't be causing it to be stuck on 0s or rather 0.0015.

I have no clue, i just thought, that this will resolve itself as soon as I'm using more iterations on my loop. As soon as I clean up my code I want to do 50 iterations instead of 10.

[Friedemannn]

I've figured it out, I was using the wrong bounds on my optimize_acqf and didnt unnormalize the candidate before giving it to my simulation package.

This fixed it:

candidates, _ = optimize_acqf(
    acq_function=acq_func,
    bounds=torch.tensor([[0.0], [1.0]], dtype=torch.double),  # normalized bounds!
    q=BATCH_SIZE,
    num_restarts=NUM_RESTARTS,
    raw_samples=RAW_SAMPLES,
)
# Unnormalize for real-world use
new_x_train = unnormalize(candidates, boundszfT).detach()
params['laser_params']['zf'] = float(new_x_train)

Now that it works I have the same spikes as with the nitrogen density one. I guess i should investigate further. Its also weird that although i used the wrong bounds and didnt unnormalize in the nitrogen density one, it still kinda worked there.

[Balandat]

Glad you were able to figure it out.

The model fits here are rather concerning though; seems like for some reason the model ends up inferring very short lengthscales. What happens if you don't pass train_Yvar and let the model infer the noise level? I wonder if there is some issue with the scaling causing the model to think it needs to become hyperlocal.

[Friedemannn]

If I don't pass noise, I get results like this:

I guess this is preferable. I at first thought that I should use as small noise values as possible, because the simulation package i use (fbpic) should be perfectly reproducible. But ig even small variable changes can lead to big differences in simulation results, so assuming noise might be a good way.
I did a BO Loop with more iterations and zoomed in on the rather big y variations on very small variations of x at the maximum:

Maybe this is also because there is some randomness in particle initialization or smth else within fbpic and i would have to fix some seeds to always get consistent results. Have to take a closer look at fbpic i guess.

But thank you, @Balandat, you already helped me a lot!

[Balandat]

Great to see this works. Yeah setting the noise to zero is dangerous in many real world use cases. Glad this is working for you now.