A physics-informed optimization algorithm

What is it about?

For best results, there is a need to fine tune the design space in computational methods that are based on machine learning models. This helps improve the model's learning process. The learning process is controlled by selecting certain values or parameters. In this paper, scientists present a hybrid optimization/exploration algorithm to select optimal parameters for models.

Featured Image

Photo by charlesdeluvio on Unsplash

Photo by charlesdeluvio on Unsplash

Why is it important?

There are multiple methods to select the right parameters for a model. In the grid search method, each value is tested individually. Next, suitable values are selected by assessing the performance of the model. But this method can take a long time to complete. That depends on the number of parameters involved. The Bayesian optimization (BO) based on a Gaussian process (GP) was introduced as a faster alternative. In it, a model learns to select the best values based on previous iterations. But this method cannot be used when the system’s behavior changes in different regions of the parameter space. The method introduced in this study combines the standard GP with a probabilistic model of the expected system’s behavior. This approach helps to make informed decisions about the points in the parameter space that must be evaluated next. This allows the algorithm to optimize the parameters by capturing trends in the data. The scientists found this algorithm to be more effective than the standard GP-BO method in optimizing parameters for 1D and 2D Ising models in physics. KEY TAKEAWAY: The algorithm introduced in the study improves the optimization process for machine learning. It can be used to learn more about a system’s properties. The approach can be used to find suitable parameters for systems that are not much known.