Model Predictive Control Utilizing Machine Learning Models within a Pinball-Based, Cyber-Physical Testbed

We examine the shot aiming problem within a physical pinball machine using model predictive control methods and machine learning based system models. A switched mode system model is developed and trained using data collected from an infrared beam-break sensor array that allows the estimation and prediction of future ball states. The trained model is then used within a model predictive controller to successfully aim shots within a physical pinball machine. The experimental results show that the controller performs with sufficient accuracy to hit standard pinball targets found in commercial pinball machines.

I. Introduction

There have been significant advances in the use of artificial intelligence (AI) and machine learning (ML) methods, in particular, deep-learning architectures [1], to develop models for applications with complex and high-dimensional data such as speech recognition [2], [3], biomedicine [4], and many other applications [5], [6]. Reinforcement learning techniques have also been used to develop automatic controllers that exhibit human-level performance in video games [7], navi-gation controllers within virtual environments [8], and many robotic applications [9]. Often these techniques represent end-to-end methods in which raw sensory information (such as pixel values) or state estimates are used to directly generate control outputs. Although these learning techniques are very effective for applications like those referenced above, they remain difficult to apply to the control of other classes of complex, real-world, cyber-physical systems.