2024/12/13 by Shah Rukh Qasim, P. Owen, Qasim, Shah Rukh +3
Computer Science · #Evolutionary Algorithms and Applications
paper · pdf · doi:10.48550/arxiv.2412.10237
We present a case for the use of Reinforcement Learning (RL) for the design\nof physics instrument as an alternative to gradient-based\ninstrument-optimization methods. It's applicability is demonstrated using two\nempirical studies. One is longitudinal segmentation of calorimeters and the\nsecond is both transverse segmentation as well longitudinal placement of\ntrackers in a spectrometer. Based on these experiments, we propose an\nalternative approach that offers unique advantages over differentiable\nprogramming and surrogate-based differentiable design optimization methods.\nFirst, Reinforcement Learning (RL) algorithms possess inherent exploratory\ncapabilities, which help mitigate the risk of convergence to local optima.\nSecond, this approach eliminates the necessity of constraining the design to a\npredefined detector model with fixed parameters. Instead, it allows for the\nflexible placement of a variable number of detector components and facilitates\ndiscrete decision-making. We then discuss the road map of how this idea can be\nextended into designing very complex instruments. The presented study sets the\nstage for a novel framework in physics instrument design, offering a scalable\nand efficient framework that can be pivotal for future projects such as the\nFuture Circular Collider (FCC), where most optimized detectors are essential\nfor exploring physics at unprecedented energy scales.\n