Skip to content

BoTorch Acquisition Functions

The BoTorch backend in ALchemist provides a flexible and powerful interface for selecting the next experiment(s) using a variety of acquisition functions from the BoTorch library. This guide explains the available options, how to use them, and what each setting means.

Overview

The Acquisition panel in ALchemist allows you to:

  • Choose between Regular, Batch, and Exploratory acquisition strategies.
  • Select from several acquisition functions, each balancing exploration and exploitation in different ways.
  • Customize parameters such as batch size and Monte Carlo integration points.
  • Run the selected strategy to suggest the next experiment(s) based on your trained model.

Important Note

You must first train your model using the BoTorch backend before running any BoTorch acquisition functions.
See BoTorch Backend for details on model training.

Acquisition Types

1. Regular Acquisition

  • Expected Improvement (EI):
    Suggests points with the highest expected improvement over the current best observed value.
  • Log Expected Improvement (LogEI):
    Numerically stable version of EI.
  • Probability of Improvement (PI):
    Selects points with the highest probability of improving over the current best value.
  • Log Probability of Improvement (LogPI):
    Numerically stable version of PI.
  • Upper Confidence Bound (UCB):
    Balances exploration and exploitation by selecting points with the highest upper confidence bound.

Customization:
- Choose to maximize or minimize your objective.

2. Batch Acquisition

  • q-Expected Improvement (qEI):
    Selects a batch of points that together maximize expected improvement.
  • q-Upper Confidence Bound (qUCB):
    Batch version of UCB.

Customization:
- Set batch size (number of points to suggest at once, q). - Monte Carlo samples (mc_samples) are used internally for batch methods.

3. Exploratory Acquisition

  • Integrated Posterior Variance (qNIPV):
    Selects points to reduce overall model uncertainty, focusing on exploration rather than optimization.

Customization:
- Set the number of Monte Carlo integration points (higher values improve accuracy but increase computation time; 500–2000 is typical).

How to Use

  1. Train Model:
    Train your model using the BoTorch backend. See BoTorch Backend for instructions.

  2. Open Acquisition Panel:
    Go to the Acquisition panel. The BoTorch options will appear automatically.

  3. Choose Acquisition Type:
    Use the segmented button to select Regular, Batch, or Exploratory.

  4. Configure Options:

  5. Select the acquisition function from the dropdown.
  6. Adjust parameters (batch size, MC points) as needed.

  7. Choose whether to maximize or minimize.

  8. Run Acquisition:
    Click Run Acquisition Strategy to suggest the next experiment(s). Results, including predicted value and uncertainty, will be shown in a notification window and highlighted in the data table and plots.

Model Optimum Finder

In addition to acquisition functions, you can use the Model Prediction Optimum tool to find the point where the model predicts the best value (maximum or minimum).
Note: This does not balance exploration and exploitation—it simply finds the model's optimum prediction.

Tips & Notes

  • Batch Acquisition: Use batch mode to suggest multiple experiments at once, useful for parallel experimentation.
  • Exploratory Mode: Use qNIPV when you want to reduce model uncertainty rather than optimize the objective.
  • Parameter Tuning: Increase MC points for more accurate but slower exploratory acquisition.
  • Publication Quality: All results and suggested points are integrated with ALchemist's visualization tools for easy analysis and export.

For more details on the underlying algorithms, see the BoTorch documentation