Search

Model-based reinforcement learning (MBRL) seeks to enhance data efficiency by learning a model of the environment and generating synthetic rollouts from it. However, accumulated model errors during these rollouts can distort the data distribution, negatively impacting policy learning and hindering long-term planning. Thus, the accumulation of model errors is a key bottleneck in current MBRL methods. We propose Infoprop, a model-based rollout mechanism that separates aleatoric from epistemic model uncertainty and reduces the influence of the latter on the data distribution. Further, Infoprop keeps track of accumulated model errors along a model rollout and provides termination criteria to limit data corruption. We demonstrate the capabilities of Infoprop in the Infoprop-Dyna algorithm, reporting state-of-the-art performance in Dyna-style MBRL on common MuJoCo benchmark tasks while substantially increasing rollout length and data quality.

Foundation models have demonstrated remarkable generalization, data efficiency, and robustness properties across various domains. In this paper, we explore the feasibility of foundation models for applications in the control domain. The success of these models is enabled by large-scale pretaining on Internet-scale datasets. These are available in fields like natural language processing and computer vision, but do not exist for dynamical systems. We address this challenge by pretraining a transformer-based foundation model exclusively on synthetic data and propose to sample dynamics functions from a reproducing kernel Hilbert space. Our pretrained model generalizes for prediction tasks across different dynamical systems, which we validate in simulation and hardware experiments, including cart-pole and Furuta pendulum setups. Additionally, the model can be fine-tuned effectively to new systems to increase performance even further. Our results demonstrate the feasibility of foundation models for dynamical systems that outperform specialist models in terms of generalization, data efficiency, and robustness., Comment: 10 pages

The consistency of a learning method is usually established under the assumption that the observations are a realization of an independent and identically distributed (i.i.d.) or mixing process. Yet, kernel methods such as support vector machines (SVMs), Gaussian processes, or conditional kernel mean embeddings (CKMEs) all give excellent performance under sampling schemes that are obviously non-i.i.d., such as when data comes from a dynamical system. We propose the new notion of empirical weak convergence (EWC) as a general assumption explaining such phenomena for kernel methods. It assumes the existence of a random asymptotic data distribution and is a strict weakening of previous assumptions in the field. Our main results then establish consistency of SVMs, kernel mean embeddings, and general Hilbert-space valued empirical expectations with EWC data. Our analysis holds for both finite- and infinite-dimensional outputs, as we extend classical results of statistical learning to the latter case. In particular, it is also applicable to CKMEs. Overall, our results open new classes of processes to statistical learning and can serve as a foundation for a theory of learning beyond i.i.d. and mixing., Comment: 26 pages, 1 figure

Dyna-style model-based reinforcement learning (MBRL) combines model-free agents with predictive transition models through model-based rollouts. This combination raises a critical question: 'When to trust your model?'; i.e., which rollout length results in the model providing useful data? Janner et al. (2019) address this question by gradually increasing rollout lengths throughout the training. While theoretically tempting, uniform model accuracy is a fallacy that collapses at the latest when extrapolating. Instead, we propose asking the question 'Where to trust your model?'. Using inherent model uncertainty to consider local accuracy, we obtain the Model-Based Actor-Critic with Uncertainty-Aware Rollout Adaption (MACURA) algorithm. We propose an easy-to-tune rollout mechanism and demonstrate substantial improvements in data efficiency and performance compared to state-of-the-art deep MBRL methods on the MuJoCo benchmark.

The industry 4.0 is leveraging digital technologies and machine learning techniques to connect and optimize manufacturing processes. Central to this idea is the ability to transform raw data into human understandable knowledge for reliable data-driven decision-making. Convolutional Neural Networks (CNNs) have been instrumental in processing image data, yet, their ``black box'' nature complicates the understanding of their prediction process. In this context, recent advances in the field of eXplainable Artificial Intelligence (XAI) have proposed the extraction and localization of concepts, or which visual cues intervene on the prediction process of CNNs. This paper tackles the application of concept extraction (CE) methods to industry 4.0 scenarios. To this end, we modify a recently developed technique, ``Extracting Concepts with Local Aggregated Descriptors'' (ECLAD), improving its scalability. Specifically, we propose a novel procedure for calculating concept importance, utilizing a wrapper function designed for CNNs. This process is aimed at decreasing the number of times each image needs to be evaluated. Subsequently, we demonstrate the potential of CE methods, by applying them in three industrial use cases. We selected three representative use cases in the context of quality control for material design (tailored textiles), manufacturing (carbon fiber reinforcement), and maintenance (photovoltaic module inspection). In these examples, CE was able to successfully extract and locate concepts directly related to each task. This is, the visual cues related to each concept, coincided with what human experts would use to perform the task themselves, even when the visual cues were entangled between multiple classes. Through empirical results, we show that CE can be applied for understanding CNNs in an industrial context, giving useful insights that can relate to domain knowledge.

Machine learning and deep learning have been used extensively to classify physical surfaces through images and time-series contact data. However, these methods rely on human expertise and entail the time-consuming processes of data and parameter tuning. To overcome these challenges, we propose an easily implemented framework that can directly handle heterogeneous data sources for classification tasks. Our data-versus-data approach automatically quantifies distinctive differences in distributions in a high-dimensional space via kernel two-sample testing between two sets extracted from multimodal data (e.g., images, sounds, haptic signals). We demonstrate the effectiveness of our technique by benchmarking against expertly engineered classifiers for visual-audio-haptic surface recognition due to the industrial relevance, difficulty, and competitive baselines of this application; ablation studies confirm the utility of key components of our pipeline. As shown in our open-source code, we achieve 97.2% accuracy on a standard multi-user dataset with 108 surface classes, outperforming the state-of-the-art machine-learning algorithm by 6% on a more difficult version of the task. The fact that our classifier obtains this performance with minimal data processing in the standard algorithm setting reinforces the powerful nature of kernel methods for learning to recognize complex patterns.

Distinguishability and, by extension, observability are key properties of dynamical systems. Establishing these properties is challenging, especially when no analytical model is available and they are to be inferred directly from measurement data. The presence of noise further complicates this analysis, as standard notions of distinguishability are tailored to deterministic systems. We build on distributional distinguishability, which extends the deterministic notion by comparing distributions of outputs of stochastic systems. We first show that both concepts are equivalent for a class of systems that includes linear systems. We then present a method to assess and quantify distributional distinguishability from output data. Specifically, our quantification measures how much data is required to tell apart two initial states, inducing a continuous spectrum of distinguishability. We propose a statistical test to determine a threshold above which two states can be considered distinguishable with high confidence. We illustrate these tools by computing distinguishability maps over the state space in simulation, then leverage the test to compare sensor configurations on hardware., Comment: 9 pages, 3 figures

We consider the problem of sequentially optimizing a time-varying objective function using time-varying Bayesian optimization (TVBO). Current approaches to TVBO require prior knowledge of a constant rate of change to cope with stale data arising from time variations. However, in practice, the rate of change is usually unknown. We propose an event-triggered algorithm, ET-GP-UCB, that treats the optimization problem as static until it detects changes in the objective function and then resets the dataset. This allows the algorithm to adapt online to realized temporal changes without the need for exact prior knowledge. The event trigger is based on probabilistic uniform error bounds used in Gaussian process regression. We derive regret bounds for adaptive resets without exact prior knowledge of the temporal changes and show in numerical experiments that ET-GP-UCB outperforms competing GP-UCB algorithms on both synthetic and real-world data. The results demonstrate that ET-GP-UCB is readily applicable without extensive hyperparameter tuning., Comment: Published in Transactions on Machine Learning Research (TMLR)

Robust controllers ensure stability in feedback loops designed under uncertainty but at the cost of performance. Model uncertainty in time-invariant systems can be reduced by recently proposed learning-based methods, which improve the performance of robust controllers using data. However, in practice, many systems also exhibit uncertainty in the form of changes over time, e.g., due to weight shifts or wear and tear, leading to decreased performance or instability of the learning-based controller. We propose an event-triggered learning algorithm that decides when to learn in the face of uncertainty in the LQR problem with rare or slow changes. Our key idea is to switch between robust and learned controllers. For learning, we first approximate the optimal length of the learning phase via Monte-Carlo estimations using a probabilistic model. We then design a statistical test for uncertain systems based on the moment-generating function of the LQR cost. The test detects changes in the system under control and triggers re-learning when control performance deteriorates due to system changes. We demonstrate improved performance over a robust controller baseline in a numerical example., Comment: To appear in the proceedings of the 61st IEEE Conference on Decision and Control

Model-based algorithms are deeply rooted in modern control and systems theory. However, they usually come with a critical assumption - access to an accurate model of the system. In practice, models are far from perfect. Even precisely tuned estimates of unknown parameters will deteriorate over time. Therefore, it is essential to detect the change to avoid suboptimal or even dangerous behavior of a control system. We propose to combine statistical tests with dedicated parameter filters that track unknown system parameters from state data. These filters yield point estimates of the unknown parameters and, further, an inherent notion of uncertainty. When the point estimate leaves the confidence region, we trigger active learning experiments. We update models only after enforcing a sufficiently small uncertainty in the filter. Thus, models are only updated when necessary and statistically significant while ensuring guaranteed improvement, which we call event-triggered learning. We validate the proposed method in numerical simulations of a DC motor in combination with model predictive control., Comment: 12 pages, 8 figures, submitted to IEEE Transactions on Control Systems Technology

Safety constraints and optimality are important, but sometimes conflicting criteria for controllers. Although these criteria are often solved separately with different tools to maintain formal guarantees, it is also common practice in reinforcement learning to simply modify reward functions by penalizing failures, with the penalty treated as a mere heuristic. We rigorously examine the relationship of both safety and optimality to penalties, and formalize sufficient conditions for safe value functions (SVFs): value functions that are both optimal for a given task, and enforce safety constraints. We reveal this structure by examining when rewards preserve viability under optimal control, and show that there always exists a finite penalty that induces a safe value function. This penalty is not unique, but upper-unbounded: larger penalties do not harm optimality. Although it is often not possible to compute the minimum required penalty, we reveal clear structure of how the penalty, rewards, discount factor, and dynamics interact. This insight suggests practical, theory-guided heuristics to design reward functions for control problems where safety is important., Comment: 16 pages, 6 figures. Accepted for publication in: Transactions of Automatic Control, special issue on Learning and Control

Most physical processes posses structural properties such as constant energies, volumes, and other invariants over time. When learning models of such dynamical systems, it is critical to respect these invariants to ensure accurate predictions and physically meaningful behavior. Strikingly, state-of-the-art methods in Gaussian process (GP) dynamics model learning are not addressing this issue. On the other hand, classical numerical integrators are specifically designed to preserve these crucial properties through time. We propose to combine the advantages of GPs as function approximators with structure preserving numerical integrators for dynamical systems, such as Runge-Kutta methods. These integrators assume access to the ground truth dynamics and require evaluations of intermediate and future time steps that are unknown in a learning-based scenario. This makes direct inference of the GP dynamics, with embedded numerical scheme, intractable. Our key technical contribution is the evaluation of the implicitly defined Runge-Kutta transition probability. In a nutshell, we introduce an implicit layer for GP regression, which is embedded into a variational inference-based model learning scheme.

Mathematical models are fundamental building blocks in the design of dynamical control systems. As control systems are becoming increasingly complex and networked, approaches for obtaining such models based on first principles reach their limits. Data-driven methods provide an alternative. However, without structural knowledge, these methods are prone to finding spurious correlations in the training data, which can hamper generalization capabilities of the obtained models. This can significantly lower control and prediction performance when the system is exposed to unknown situations. A preceding causal identification can prevent this pitfall. In this paper, we propose a method that identifies the causal structure of control systems. We design experiments based on the concept of controllability, which provides a systematic way to compute input trajectories that steer the system to specific regions in its state space. We then analyze the resulting data leveraging powerful techniques from causal inference and extend them to control systems. Further, we derive conditions that guarantee the discovery of the true causal structure of the system. Experiments on a robot arm demonstrate reliable causal identification from real-world data and enhanced generalization capabilities., Comment: Accepted final versions to appear in the Transactions on Machine Learning Research

Evaluating whether data streams are drawn from the same distribution is at the heart of various machine learning problems. This is particularly relevant for data generated by dynamical systems since such systems are essential for many real-world processes in biomedical, economic, or engineering systems. While kernel two-sample tests are powerful for comparing independent and identically distributed random variables, no established method exists for comparing dynamical systems. The main problem is the inherently violated independence assumption. We propose a two-sample test for dynamical systems by addressing three core challenges: we (i) introduce a novel notion of mixing that captures autocorrelations in a relevant metric, (ii) propose an efficient way to estimate the speed of mixing relying purely on data, and (iii) integrate these into established kernel two-sample tests. The result is a data-driven method that is straightforward to use in practice and comes with sound theoretical guarantees. In an example application to anomaly detection from human walking data, we show that the test is readily applicable without any human expert knowledge and feature engineering.

Despite the availability of ever more data enabled through modern sensor and computer technology, it still remains an open problem to learn dynamical systems in a sample-efficient way. We propose active learning strategies that leverage information-theoretical properties arising naturally during Gaussian process regression, while respecting constraints on the sampling process imposed by the system dynamics. Sample points are selected in regions with high uncertainty, leading to exploratory behavior and data-efficient training of the model. All results are finally verified in an extensive numerical benchmark.

When models are inaccurate, the performance of model-based control will degrade. For linear quadratic control, an event-triggered learning framework is proposed that automatically detects inaccurate models and triggers the learning of a new process model when needed. This is achieved by analyzing the probability distribution of the linear quadratic cost and designing a learning trigger that leverages Chernoff bounds. In particular, whenever empirically observed cost signals are located outside the derived confidence intervals, we can provably guarantee that this is with high probability due to a model mismatch. With the aid of numerical and hardware experiments, we demonstrate that the proposed bounds are tight and that the event-triggered learning algorithm effectively distinguishes between inaccurate models and probabilistic effects such as process noise. Thus, a structured approach is obtained that decides when model learning is beneficial., Comment: 13 pages, 8 figures, accepted for publication in IEEE Transactions on Automatic Control

Communication load is a limiting factor in many real-time systems. Event-triggered state estimation and event-triggered learning methods reduce network communication by sending information only when it cannot be adequately predicted based on previously transmitted data. This paper proposes an event-triggered learning approach for nonlinear discrete-time systems with cyclic excitation. The method automatically recognizes cyclic patterns in data - even when they change repeatedly - and reduces communication load whenever the current data can be accurately predicted from previous cycles. Nonetheless, a bounded error between original and received signal is guaranteed. The cyclic excitation model, which is used for predictions, is updated hierarchically, i.e., a full model update is only performed if updating a small number of model parameters is not sufficient. A nonparametric statistical test enforces that model updates happen only if the cyclic excitation changed with high probability. The effectiveness of the proposed methods is demonstrated using the application example of wireless real-time pitch angle measurements of a human foot in a feedback-controlled neuroprosthesis. The experimental results show that communication load can be reduced by 70 % while the root-mean-square error between measured and received angle is less than 1{\deg}., Comment: 6 pages and 6 figures; to appear in IEEE Control Systems Letters

The efficient exchange of information is an essential aspect of intelligent collective behavior. Event-triggered control and estimation achieve some efficiency by replacing continuous data exchange between agents with intermittent, or event-triggered communication. Typically, model-based predictions are used at times of no data transmission, and updates are sent only when the prediction error grows too large. The effectiveness in reducing communication thus strongly depends on the quality of the prediction model. In this article, we propose event-triggered learning as a novel concept to reduce communication even further and to also adapt to changing dynamics. By monitoring the actual communication rate and comparing it to the one that is induced by the model, we detect a mismatch between model and reality and trigger model learning when needed. Specifically, for linear Gaussian dynamics, we derive different classes of learning triggers solely based on a statistical analysis of inter-communication times and formally prove their effectiveness with the aid of concentration inequalities.

In networked control systems, communication is a shared and therefore scarce resource. Event-triggered control (ETC) can achieve high performance control with a significantly reduced amount of samples compared to classical, periodic control schemes. However, ETC methods usually rely on the availability of an accurate dynamics model, which is oftentimes not readily available. In this paper, we propose a novel event-triggered pulse control strategy that learns dynamics models if necessary. In addition to adapting to changing dynamics, the method also represents a suitable replacement for the integral part typically used in periodic control., Comment: Accepted final version to appear in: Proc. of the American Control Conference, 2019

Encoding a sequence of observations is an essential task with many applications. The encoding can become highly efficient when the observations are generated by a dynamical system. A dynamical system imposes regularities on the observations that can be leveraged to achieve a more efficient code. We propose a method to encode a given or learned dynamical system. Apart from its application for encoding a sequence of observations, we propose to use the compression achieved by this encoding as a criterion for model selection. Given a dataset, different learning algorithms result in different models. But not all learned models are equally good. We show that the proposed encoding approach can be used to choose the learned model which is closer to the true underlying dynamics. We provide experiments for both encoding and model selection, and theoretical results that shed light on why the approach works.

An efficient representation of observed data has many benefits in various domains of engineering and science. Representing static data sets, such as images, is a living branch in machine learning and eases downstream tasks, such as classification, regression, or decision making. However, the representation of dynamical systems has received less attention. In this work, we develop a method to represent a dynamical system efficiently as a combination of a state and a local model, which fulfills a criterion inspired by the minimum description length (MDL) principle. The MDL principle is used in machine learning and statistics to quantify the trade-off between the ability to explain seen data and the model complexity. Networked control systems are a prominent example, where such a representation is beneficial. When many agents share a network, information exchange is costly and should thus happen only when necessary. We empirically show the efficiency of the proposed encoding for several dynamical systems and demonstrate reduced communication for event-triggered state estimation problems., Comment: 7 pages, 5 figures, to appear in 57th IEEE Conference on Decision and Control (CDC 2018)

Common event-triggered state estimation (ETSE) algorithms save communication in networked control systems by predicting agents' behavior, and transmitting updates only when the predictions deviate significantly. The effectiveness in reducing communication thus heavily depends on the quality of the dynamics models used to predict the agents' states or measurements. Event-triggered learning is proposed herein as a novel concept to further reduce communication: whenever poor communication performance is detected, an identification experiment is triggered and an improved prediction model learned from data. Effective learning triggers are obtained by comparing the actual communication rate with the one that is expected based on the current model. By analyzing statistical properties of the inter-communication times and leveraging powerful convergence results, the proposed trigger is proven to limit learning experiments to the necessary instants. Numerical and physical experiments demonstrate that event-triggered learning improves robustness toward changing environments and yields lower communication rates than common ETSE., Comment: 7 pages, 4 figures, to appear in the 2018 American Control Conference (ACC)