Modeling Deep Reinforcement Learning based Architectures for Cyber-Physical Systems

Reinforcement learning is a sub-field of machine learning where an agent aims to learn a behavior or a policy maximizing a reward function by trial and error. The approach is particularly interesting for the design of autonomous cyber-physical systems such as self-driving cars. In this work we present a generative, domain-specific modeling framework for the design, training and integration of reinforcement learning systems. It consists of a neural network modeling language which is used to design the models to be trained, e.g. actor and critic networks, and a training language used to describe the training procedure and set the corresponding hyperparameters. The underlying component model allows the modeler to embed the trained networks in larger component & connector architectures. We illustrate our framework by the example of a self-driving racing car.