Improving Large Distributed Control System Development and Dependability: An Approach using New Simulation Structures and Reinforcement Learning Procedures
August 9, 2019
Light Engineering room 250
Advisor: Thomas G. Robertazzi
The Relativistic Heavy Ion Collider (RHIC) is a world-class particle accelerator at Brookhaven National Laboratory (BNL). It enables scientists to study what the universe may have looked like in the first few moments after its creation. RHIC contains two 3.8 kilometers counter-rotating super-conducting rings to carry particle beams which can be collided in six crossing regions to provide possible interactions for experimenters to study.
The control system of the Collider-Accelerator Department (C-AD) at BNL is a large distributed system. It provides operational interfaces to the collider and injection beam lines. It has approximately 1.5 million control points. Instances of the C-AD control systems are applied in the Linear Accelerator (Linac), Electron Beam Ion Source (EBIS), Tandem Van de Graff pre-accelerators, the Booster accelerator, Alternating Gradient Synchrotron (AGS), and the RHIC. Its performance has a crucial impact over the whole accelerator suite.
This work aims to improve the robustness of the C-AD control systems. It consists of two parts. The first part introduces a simulation framework, which aims to improve the software codes reliability in the control system. The second part analyzes a fundamental performance bottleneck in the control system. Several reinforcement learning based algorithms are proposed to optimize various goals.