Exploration of different parameter optimization algorithms within the context of ACTS software framework

R. B. Garg, H. M. Gray, E. Hofgard, L. TompkinsResearch Square 2023 · DOI

Abstract

Particle track reconstruction, in which the trajectories of charged particles are determined, is a critical and time consuming component of the full event reconstruction chain. The underlying software is complex and consists of a number of mathematically intense algorithms, each dealing with a particular tracking sub-process. These algorithms have many input parameters that need to be supplied in advance. However, it is difficult to determine the configuration of these parameters that produces the best performance. Currently, the input parameter values are decided on the basis of prior experience or by the use of brute force techniques. A parameter optimization approach that is able to automatically tune these parameters for high performance is greatly desirable. In the current work, we explore various machine learning based optimization methods to devise a suitable technique to optimize parameters in the complex track reconstruction environment. These methods are evaluated on the basis of a metric that targets high efficiency while keeping the duplicate and fake rates small. We focus on derivative free optimization approaches that can be applied to problems involving non-differentiable loss functions. For our studies, we consider the tracking algorithms defined within A Common Tracking Software (ACTS) framework. We test our methods using simulated data from ACTS software corresponding to the ACTS Generic detector and the ATLAS ITk detector geometries.

Introduction

In high-energy physics, "tracking" is the essential process of reconstructing the trajectories of charged particles from discrete signals left in a detector. This task is computationally intensive and relies on complex software frameworks like A Common Tracking Software (ACTS). These frameworks utilize numerous internal parameters—such as search tolerances and thresholds—that significantly impact performance. Traditionally, these parameters are tuned using physicist intuition or exhaustive manual searches. However, the upcoming High-Luminosity LHC (HL-LHC) era presents unprecedented data volumes and "pile-up" conditions, making manual tuning unsustainable. This research investigates whether automated, machine-learning-based optimization can outperform traditional methods in this noisy, non-differentiable landscape.

Methods

The study integrated ACTS with five derivative-free optimization (DFO) algorithms. These "black-box" optimizers evaluate the output "score" of the tracking framework and suggest improved input parameters without requiring knowledge of the underlying physics equations.

The closed-loop parameter optimization process between the ACTS framework and the tuning algorithm.

The algorithms evaluated include:

  1. Evolutionary Algorithms (EA): Utilizing principles of natural selection to evolve optimal parameter sets.
  2. Optuna (TPE): A Bayesian approach employing Tree-structured Parzen Estimators.
  3. Lipschitz Optimization (LIPO): A mathematical method assuming a bounded rate of change in the objective function.
  4. Scikit Optimization (Skopt): Specifically the "forest_optimize" variant, which uses decision trees to model parameter space.
  5. Orion: Used as a baseline framework for simple random searches.

The optimization focused on two critical components: the Combinatorial Kalman Filter (CKF) for track finding and the Adaptive Multi Vertex Finder (AMVF) for vertex reconstruction.

Datasets and Experimental Setup

Performance was assessed using simulated top-antitop ($t\bar{t}$) quark production data, a standard benchmark for tracking software.

Cross-section of the detector geometry used for simulating particle tracks.

The study utilized two distinct detector geometries:

  • ACTS Generic Detector: A simplified silicon detector model.
  • ATLAS ITk: The planned Inner Tracker upgrade for the ATLAS experiment.

Simulations were conducted under high "pile-up" conditions (140 and 200 simultaneous collisions) to mimic the extreme environments expected at the HL-LHC.

Key Insights and Results

The automated approach demonstrated significant improvements over manually tuned "default" settings. A crucial factor was the development of a custom "Score Function" that balanced track efficiency against "fakes" (incorrect tracks), "duplicates" (redundant tracks), and computational execution time.

Performance Highlights for the CKF Algorithm: * Efficiency: Increased from a default of 93.6% to 96.8% (using LIPO). * Duplicate Rate: Reduced from 72.6% to 56.6% (using Skopt). * Processing Speed: Execution time per event was slashed from 50.2 seconds to 31.1 seconds (using Evolutionary Algorithms).

The results also indicated that the optimization landscape is "bumpy," with different algorithms finding various local minima that yielded similar high performance. Importantly, the success across both generic and specific (ATLAS ITk) geometries suggests that these ML tools are highly portable.

Conclusion and Future Work

Automated machine learning optimization successfully outperformed human-tuned defaults, offering a path toward more efficient tracking in the HL-LHC era. While this study optimized a subset of parameters (8 for CKF and 5 for AMVF), future work will expand this to dozens of parameters simultaneously. Additionally, researchers aim to reduce the computational overhead of the tuning process itself and explore real-time integration into experimental "trigger" systems to adapt to changing detector conditions dynamically.

Summary updated

Abstract

<title>Abstract</title> Particle track reconstruction, in which the trajectories of charged particles are determined, is a critical and time consuming component of the full event reconstruction chain. The underlying software is complex and consists of a number of mathematically intense algorithms, each dealing with a particular tracking sub-process. These algorithms have many input parameters that need to be supplied in advance. However, it is difficult to determine the configuration of these parameters that produces the best performance. Currently, the input parameter values are decided on the basis of prior experience or by the use of brute force techniques. A parameter optimization approach that is able to automatically tune these parameters for high performance is greatly desirable. In the current work, we explore various machine learning based optimization methods to devise a suitable technique to optimize parameters in the complex track reconstruction environment. These methods are evaluated on the basis of a metric that targets high efficiency while keeping the duplicate and fake rates small. We focus on derivative free optimization approaches that can be applied to problems involving non-differentiable loss functions. For our studies, we consider the tracking algorithms defined within A Common Tracking Software (ACTS) framework. We test our methods using simulated data from ACTS software corresponding to the ACTS Generic detector and the ATLAS ITk detector geometries.
BibTeX
@inproceedings{garg2023exploration,
  author = {R. B. Garg and H. M. Gray and E. Hofgard and L. Tompkins},
  title = {{Exploration of different parameter optimization algorithms within the context of ACTS software framework}},
  booktitle = {Research Square},
  year = {2023},
  doi = {10.21203/rs.3.rs-2729443/v1},
}
Figure