qNoisyExpectedImprovement is an expected-improvement acquisition policy for Bayesian optimization when observations are noisy or the surrogate model remains substantially uncertain, including at observed training locations. section 10.4
It follows a “king of the hill” strategy: seek a candidate expected to improve on the best point found so far, then compare the incumbent and challenger. section 10.4 section 10.5
A typical optimization step is:
- Fit or update the Gaussian-process model with the available preference data. section 10.4
- Construct qNoisyExpectedImprovement with the model and training inputs, then optimize it over the search bounds. section 10.4
- Place the recommended challenger against the current best point and append the resulting pairwise feedback to the training set. section 10.4 section 10.5
Preference optimization
For a user preference function observed through pairwise comparisons, select a challenger with qNoisyExpectedImprovement, compare it with the incumbent, and use the preference outcome to update the optimization data. section 10.4
policy = qNoisyExpectedImprovement(model, train_x) #1
challenger, acq_val = optimize_acqf( #2
policy, #2
bounds=bounds, #2
q=1, #2
num_restarts=50, #2
raw_samples=100, #2
) #2
The listing shows the BoTorch construction of qNoisyExpectedImprovement and its use with optimize_acqf, bounded search, restarts, and raw samples. section 10.4
