Models for data-efficient reinforcement learning on real-world applications

Abstract

Large-scale deep Reinforcement Learning is strongly contributing to many recently published success stories of Artificial Intelligence. These techniques enabled computer systems to autonomously learn and master challenging problems, such as playing the game of Go or complex strategy games such as Star-Craft on human levels or above. Naturally, the question arises which problems could be addressed with these Reinforcement Learning technologies in industrial applications. So far, machine learning technologies based on (semi-)supervised learning create the most visible impact in industrial applications. For example, image, video or text understanding are primarily dominated by models trained and derived autonomously from large-scale data sets with modern (deep) machine learning methods. Reinforcement Learning, on the opposite side, however, deals with temporal decision-making problems and is much less commonly found in the industrial context. In these problems, current decisions and actions inevitably influence the outcome and success of a process much further down the road. This work strives to address some of the core problems, which prevent the effective use of Reinforcement Learning in industrial settings. Autonomous learning of new skills is always guided by existing priors that allow for generalization from previous experience. In some scenarios, non-existing or uninformative prior knowledge can be mitigated by vast amounts of experience for a particular task at hand. Typical industrial processes are, however, operated in very restricted, tightly calibrated operating points. Exploring the space of possible actions or changes to the process naively on the search for improved performance tends to be costly or even prohibitively dangerous. Therefore, one reoccurring subject throughout this work is the emergence of priors and model structures that allow for efficient use of all available experience data. A promising direction is Model-Based Reinforcement Learning, which is explored in the first part of this work. This part derives an automatic tuning method for one of themostcommonindustrial control architectures, the PID controller. By leveraging all available data about the system’s behavior in learning a system dynamics model, the derived method can efficiently tune these controllers from scratch. Although we can easily incorporate all data into dynamics models, real systems expose additional problems to the dynamics modeling and learning task. Characteristics such as non-Gaussian noise, latent states, feedback control or non-i.i.d. data regularly prevent using off-the-shelf modeling tools. Therefore, the second part of this work is concerned with the derivation of modeling solutions that are particularly suited for the reinforcement learning problem. Despite the predominant focus on model-based reinforcement learning as a promising, data-efficient learning tool, this work’s final part revisits model assumptions in a separate branch of reinforcement learning algorithms. Again, generalization and, therefore, efficient learning in model-based methods is primarily driven by the incorporated model assumptions (e.g., smooth dynamics), which real, discontinuous processes might heavily violate. To this end, a model-free reinforcement learning is presented that carefully reintroduces prior model structure to facilitate efficient learning without the need for strong dynamic model priors. The methods and solutions proposed in this work are grounded in the challenges experienced when operating with real-world hardware systems. With applications on a humanoid upper-body robot or an autonomous model race car, the proposed methods are demonstrated to successfully model and master their complex behavior.