Particle Identification

A Comparative Analysis of Boosting and Artificial Neural Networks for Particle Identification at ALICE.



Background

A Large Ion Collider Experiment (ALICE) is the dedicated heavy ion experiment at the Large Hadron Collider (LHC). ALICE collides heavy ions - the nuclei of heavy particles - at nearly the speed of light. In this paper we consider how to identify the particles that are emitted from these collisions, rather like subatomic shrapnel. ALICE has been operating since 2010 and it is scheduled for an upgrade in 2017. Most importantly, the upgrade will increase the rate and momentum of particle collisions improving the quality of results for current experimentation and allowing for new experimentation. The new system will also impose new hardware requirements. As a result the current particle identification framework needs to be redeveloped.

Currently, artificial neural networks are used for particle identification at ALICE. A similar experiment, MiniBooNE at Fermilabs, uses boosting - a technique of combining many weak classifiers into a single strong classifier - to identify particles. After the upgrade the new hard- ware at ALICE will require that the information used for particle identification gets compressed into six 8-bit values, which is currently not necessary at either ALICE or Mini- BooNE. Consequently, the systems now in use at ALICE and MiniBooNE will not be compatible.

Aim

ALICE’s upgrade is imposing new hardware limitations that require the development of a new particle identification algorithm. This means there is a need to determine what type of machine learning algorithm would be suited to the type of problem presented by ALICE’s particle identification requirements after the upgrade.

This project aims to determine whether boosting or artificial neural networks are better suited for particle identification at ALICE after the upgrade. A related aim is to establish which algorithms and parameters within these two large classes of algorithms perform well.

Approach

To compare boosting and artificial neural networks, a variety of adaptive boosting and feed-forward artificial neural network algorithms are implemented. Adaptive boosting and feed-forward artificial neural networks are the base on which the boosting and artificial neural network class of algorithms is built. This means that the algorithms implemented in this project will give an indication of how well each class of algorithms would perform at ALICE after the upgrade. The algorithms are tested on simulated data generated by AliRoot - ALICE’s software package.

Three boosting algorithms, AdaBoost ε-Boost and ε-LogitBoost, were implemented and tested. These three boosting algorithms performed well at MiniBooNE a similar particle identification experiment and therefore were good candidates. Test were done using decision trees and artificial neural networks as weak learners.

Feed forward artificial neural networks using the back propagation algorithm to update the weights with various hidden layer topologies and activation functions where also tested.

From these boosting and artificial neural network algorithms tested the best performing models where compared to determine which would be better suited to ALICE after the upgrade.

Results

In this project various boosting models and artificial neural networks have been tested on simulated data to determine which will be better suited to the upgraded system at ALICE.

From the set of boosting algorithms tested it was found that ε-LogitBoost was the worst performing. ε-Boost performed better than Adaboost on complex weak learners but worse than AdaBoost on simple weak learners. Since using simple weak learners generalises better AdaBoost was determined to be the best performing boosting algorithm.

Using artificial neural networks with two or three nodes as weak learners performed better than decision trees. These small artificial neural networks learnt slower and continued to improve performance for longer than the other weak learners tested.

As a result the best performing boosting algorithm was determined to be AdaBoost using artificial neural networks with two nodes in the hidden layer.

Artificial neural networks using the sigmoid activation function learnt the data better than the fast-sigmoid activation function and over fitted the data less than the tanh activation function. artificial neural networks with a single hidden layer performed best. As a result it was determined that artificial neural networks with a single hidden layer using the sigmoid activation function was the best performing artificial neural network.

The best performing artificial neural network - single hidden layer and sigmoid activation function - was compared to the best performing boosting algorithm - AdaBoost using artificial neural networks with two nodes in the hidden layer. The analysis shows that the artificial neural network was more accurate and computationally more efficient in terms of training and classifying.

Figure 1. Figure showing the performance of the best performing boosting algorithm - AdaBoost.
Figure 2. Figure showing the performance of the best performing artificial neural network.
Figure 3. Figure comparing the performance of the three different boosting algorithms.

Conclusions

Considered collectively, our experiments indicate that after the upgrade at ALICE artificial neural networks should be used for particle identification.

From the experiments done in this project it became apparent that boosting has performance benefits over artificial neural neworks when the number of predictors is large. This is because boosting has a simpler weighting scheme making it easier to find an optimal solution. At ALICE there are only six predictors per particle. This nullifies the benefit that boosting algorithms have over artificial neural networks.