The application of machine learning techniques toward problems in the physical sciences is really taking off!
Here’s an article about another group that used decision theory (similar to what we’ve developed at CASTLE Labs) to replicated a Nobel Prize-winning result in the field of Bose-Einstein Condensation.
They used a decision policy to determine the next experiment based on past data. This machine learning tool recommended the temperature cool-down schedule in the experiment, a crucial piece in trying to form a Bose-Einstein condensate (BEC). Over several iterations of the decide-experiment-measure main loop, the tool was able to find a temperature ramp that resulted in in BEC.
Check out the article in Nature Scientific Reports.
What’s interesting here is that the decision policy the tool used is a combination of two very simple policies: pure exploration (which picks an experiment at random) and variance reduction (which picks an experiment based on how much we don’t know a priori about the eventual result). In our lab, we’ve developed a sophisticated policy known as the Knowledge Gradient (KG) policy, which typically out-performs these simpler policies. What this means is that, using the KG policy, we can reduce the number of experiments needed in order to optimize any objective.