One of the goals of the BaBar experiment at the Stanford Linear Accelerator Center (SLAC) is to examine the difference between matter and anti-matter. This involves experiments where the collision of beams generates events that produce a B particle and its anti-particle B-bar. After identifying a B/B-bar event, the major challenge is to classify it as either B or B-bar. Both particles live for a very short time, in the average around 10-12 seconds. Hence we can only observe their decay products and have to “reconstruct” the B or B-bar from these decay products. There are many ways the B and B-bar can decay, each decay mode happening randomly but not with equal probability. The typical probability for any one decay mode is less than 10-4, many below 10-5. In other words, there is no unique signature. Furthermore, many B decays and B-bar decays have similar signatures. For example, the momentum distribution of a decay product is systematically higher if it came from a B rather than a B-bar; however, the two momentum distributions overlap significantly. Even if we have a perfect detector with exquisite resolution and never make a mistake, we can only identify the B from the B-bar on a statistical basis. We refer to this procedure as “tagging” the B or the B-bar. There are many discriminating variables like this, each with a very modest separation. They are often correlated, but not necessarily in a straight forward way. We want to combine them and achieve the best tagging performance.
Currently we are evaluating the performance of existing non-parametric classifiers in order to determine their suitability for tagging B/B-bar events. Based on these findings we develop novel approaches that improve classification accuracy and will ultimately allow a more efficient usage of the data which is obtained from resource-intensive experimental runs. Another goal is to examine how these new techniques can be applied to other scientific data, e.g., other high-energy physics experiments.
Charles C. Young (SLAC)
Rich Caruana (Cornell)
Johannes Gehrke (Cornell)
Mirek Riedewald (Cornell)