Ehrhardt, Jonas

This is an extended abstract of the manuscript "A Lazy Approach to Neural Numerical Planning with Control Parameters". The paper presents a lazy, hierarchical approach to tackle the challenge of planning in complex numerical domains, where the effects of actions are influenced by control parameters, and may be described by neural networks.

Autonomously detecting anomalous behavior based on system observations is a fundamental task for Cyber-Physical Systems (CPS). Due to the high system complexity and large number of subsystems in modern CPS, rule- or knowledge-based approaches for anomaly detection are more and more replaced by Machine Learning (ML) approaches which leverage historical CPS data. Typically, ML approaches learn a system model based on the CPS data and identify anomalous behavior based on the distance of the real CPS behavior to the predicted model behavior. However, most classical ML approaches for anomaly detection are monolithic, meaning a single ML model is fitted on a global CPS observation, making them frail to spurious correlations and confounders that originate on CPS subsystem level. We hence propose a modular approach toward anomaly detection in CPS, specifically a novel Modular Neural Network (MNN) architecture. Our architecture not only models the behavior of individual CPS sub-systems in individual MNN modules, but additionally models the dependencies of the CPS subsystems into the MNN architecture. Thereby, we omit confounding effects and spurious correlations, enabling us to identify and allocate anomalies within the CPS at subsystem level. We benchmark our MNN architecture against monolithic Neural Networks and MNN architectures that do not explicitly model CPS subsystem dependencies using a real-world dataset of an industrial robot with different anomalies. We show that by modeling real-world dependencies into a MNN architecture, we can improve the performance of autonomous anomaly detection in CPS.

Approaches in AI planning for Cyber-Physical Production Systems (CPPS) are mainly symbolic and depend on comprehensive formalizations of system domains and planning problems. Handcrafting such formalizations requires detailed knowledge of the formalization language, of the CPPS, and is overall considered difficult, tedious, and error-prone. Within this paper, we suggest a sub-symbolic approach for solving planning problems in CPPS. Our approach relies on neural networks that learn the dynamical behavior of individual process steps from global time-series observations of the CPPS and are embedded in a superordinate network architecture. In this context, we present the process step representation network architecture (peppr), a novel neural network architecture, which can learn the behavior of individual or multiple dynamical systems from global time-series observations. We evaluate peppr on real datasets from physical and biochemical CPPS, as well as artificial datasets from electrical and mathematical domains. Our model outperforms baseline models like multilayer perceptrons and variational autoencoders and can be considered as a first step towards a sub-symbolic approach for planning in CPPS.

In this paper, we tackle the problem of planning in complex numerical domains, where actions are indexed by control parameters, and their effects may be described by neural networks. We propose a lazy, hierarchical approach based on two ingredients. First, a Satisfiability Modulo Theory solver looks for an abstract plan where the neural networks in the model are abstracted into uninterpreted functions. Then, we attempt to concretize the abstract plan by querying the neural network to determine the control parameters. If the concretization fails and no valid control parameters could be found, suitable information to refine the abstraction is lifted to the Satisfiability Modulo Theory model. We contrast our work against the state of the art in NN-enriched numerical planning, where the neural network is eagerly and exactly represented as terms in Satisfiability Modulo Theories over nonlinear real arithmetic. Our systematic evaluation on four different planning domains shows that avoiding symbolic reasoning about the neural network not only leads to substantial efficiency improvements, but also enables their integration as black-box models.