Abstract
The efficacy of particle identification is compared using artificial neutral networks and boosted decision trees. The comparison is performed in the context of the MiniBooNE, an experiment at Fermilab searching for neutrino oscillations. Based on studies of Monte Carlo samples of simulated data, particle identification with boosting algorithms has better performance than that with artificial neural networks for the MiniBooNE experiment. Although the tests in this paper were for one experiment, it is expected that boosting algorithms will find wide application in physics.
Introduction
In high-energy physics, the search for rare phenomena often requires distinguishing a tiny signal from an overwhelming background. For the MiniBooNE experiment at Fermilab, this challenge manifests as identifying rare neutrino oscillations ($\nu_\mu \to \nu_e$) amidst background noise like misidentified pions. For years, Artificial Neural Networks (ANNs) were the standard tool for this Particle Identification (PID) task. However, ANNs can be difficult to optimize, requiring extensive tuning of architectures and input scaling. This paper explores Boosted Decision Trees (BDTs) as a more robust and higher-performing alternative.
Methodology: From Networks to Trees
The study compared traditional multi-layer perceptron ANNs against BDTs using large-scale Monte Carlo simulations. The dataset included 200,000 signal events and 500,000 background events, characterized by 52 distinct PID variables such as hit multiplicity, energy, and radial position.
Decision Trees and Boosting
A decision tree is a sequence of binary splits designed to maximize the purity of the resulting branches. Unlike ANNs, which process variables through weighted sums and non-linear activation functions, decision trees use simple cuts on individual variables.
To overcome the limitations of a single "weak" classifier, the authors employed Boosting. This technique builds a sequence of many trees (often 1,000 or more). If an event is misclassified by one tree, its weight is increased ("boosted") so that subsequent trees focus more on correcting that error. The final classification score is a weighted sum of the outputs from all trees in the ensemble.
Key Results and Performance
The primary finding of the study is that BDTs consistently outperform ANNs in background rejection. Specifically, BDTs were found to be 20% to 80% more effective than standard ANN PID methods. For instance, at a 50% signal efficiency, the ANN typically retained approximately 1.5 times more background events than the BDT.
Beyond raw performance, BDTs offered several practical advantages:
- Robustness: BDTs are invariant under monotonic transformations of input variables (like taking a logarithm), whereas ANNs are highly sensitive to input scaling.
- Feature Handling: BDTs naturally handle a large number of input variables, ignoring those with little discriminating power and exploiting subtle correlations in others. This eliminates the need for the intensive manual feature selection often required for ANNs.
- Stability: The ensemble approach of boosting makes the final classifier much more stable and less prone to the optimization pitfalls of a single complex model.
Conclusion
The transition from ANNs to BDTs represents a significant step forward in event classification for high-energy physics. By turning an ensemble of simple decision trees into a powerful classifier through boosting, researchers at MiniBooNE achieved superior background rejection with less manual tuning. While this study was specific to one experiment, the robustness and efficiency of BDTs have since led to their widespread adoption across the field of particle physics.
Summary updated