Nils Rosemann

Stable Classification in Environments with Varying Degrees of Uncertainty

Most practical signal processing problems have to deal with uncertainties, e. g., due to noisy input data. Usual strategies to do this are based on estimating these uncertainties by statistical methods in advance. For some systems with multi-staged signal processing it is possible to identify these estimates at runtime and to relate a degree of certainty to them. If such degrees of certainty are known for input signals, e. g. by earlier stages of processing, this knowledge can be used to get a more robust or accurate result in classification tasks in the later stages, even if they vary at runtime. In this paper we thus introduce an approach to extend support vector machines to incorporate such known uncertainties at runtime, given as certainty degrees. Based on the known certainty of each input, classification depends more on certain inputs and gradually less on uncertain input data. This is done by changing the decision (kernel) function online, i. e., during operation. An artificial two-dimensional dataset is used to visualize the effects of this extension. And the application to three different datasets is a first benchmark showing that the resulting classification quality increases when known uncertainties are considered.

The Incremental Risk Functional: Basics of a Novel Incremental Learning Approach

Incremental learning gets increasing attention in research and practice as it has the advantages of continuous adaptation and handling big data with a low computation and memory demand at the same time. Several approaches have been proposed recently for online learning, but only few work has been done to regard the influence of the approximation structure. Hence, we introduce the incremental risk functional which directly incorporates knowledge about the approximation structure into its parameter update. Exemplary, we apply this approach to regression estimation through linear-in-parameter approximators. We show that the resulting learning algorithm converges and changes the global functional behavior only as little as necessary with every learning step, thus resulting in a stable incremental learning approach.

Concept for Controlled Self-optimization in Online Learning Neuro-fuzzy Systems

Many modern control systems, e.g., in automotive or robotic applications get increasingly complex and hard to design. This is due to the complex interactions of their internal subsystems, but additionally, these systems operate in a dynamically changing, complex environment. The Organic Computing (OC) initiative tries to cope with the resulting engineering demands by introducing emergence and self-x properties into the systems (e.g., self-organization, self-optimization). Within this context, we focus on control systems which adapt their behavior autonomously by learning.

A Framework for Controlled Self-optimisation in Modular System Architectures

Organic Computing tackles design issues of future technical systems by equipping them with self-x properties. A key self-x feature is self-optimisation, i.e. the system’s ability to adapt its dynamic behaviour to its current environment and requirements. In this article, it is shown how self-optimisation can be realised in a safe and goal-directed way, but also why it has to be enhanced and embedded into a suitable, modular system architecture. Then, a suitable framework for controlled self-optimisation is developed, which enables the system designer to give a priori guarantees of important dynamic system properties, and which ensures the system’s ability to cope dynamically with anomalies. The key features are online machine learning, which is complemented by incremental, local regularisation in a local Observer/Controller architecture as well as expressing anomalies by health signals, which are exploited to guide the learning process dynamically in order to achieve fast, but safe learning.

ORCA: An Organic Robot Control Architecture

Mastering complexity is one of the greatest challenges for future dependable information processing systems. Traditional fault tolerance techniques relying on explicit fault models seem to be not sufficient to meet this challenge. During their evolution living organisms have, however, developed very effective and efficient mechanisms like the autonomic nervous system or the immune system to make them adaptive and self-organising. Thus, they are able to cope with anomalies, faults or new unforeseen situations in a safe way. Inspired by these organic principles the control architecture ORCA (Organic Robot Control Architecture) was developed. Its aim is to transfer self-x properties from organic to robotic systems. It is described in this article with a specific focus on the way ORCA deals with dynamically changing uncertainties and anomalies.

Instantaneous anomaly detection in online learning fuzzy systems

In the field of self-optimizing automation systems, incremental local learning is an important technique. But especially in case of closed loop coupling, learnt anomalies may have a negative influence on the entire future of the evolving system. In the worst case, this may result in unstable or chaotic system behavior. Thus it is crucial to detect anomalies in online learning systems instantaneously to be able to take immediate counteractions. This paper presents an intuitive approach how to detect anomalies in incrementally and locally learning TS-fuzzy systems by looking at local meta-level characteristics of the learnt function. The practical feasibility of this approach is then investigated in experiments with a real pole-balancing cart.

Trust Management—Handling Uncertainties in Embedded Systems

This article summarises the current status of the Trust-Management project and gives an outlook to further research.

Modellierung von Anomalien in einer modularen Roboter-Steuerung

Moderne Robotersysteme werden immer komplexer und dadurch schwieriger zu entwerfen. Auserdem steigt die Wahrscheinlichkeit von Fehlern. Organic Computing versucht durch Methoden organischer Systeme wie Emergenz und Selbstorganisation gleichzeitig das Entwurfsproblem zu losen und das autonome Reagieren auf Fehler zu erreichen, ohne den Aufwand der klassischen Fehlertoleranz zu investieren. Um auf anormale Situationen reagieren zu konnen, muss quasi der „Gesundheitszustand“ des Roboters erfasst werden. In diesem Beitrag wird beschrieben, wie dieser ausgehend von Sensorsignalen auf den verschiedenen Ebenen einer Steuerungshierarchie reprasentiert werden kann. Ein wesentlicher Mechanismus dazu sind „Health-Signale“. Ihre Semantik und systematische Verknupfung werden naher erlautert und an einem einfachen Beispiel demonstriert.

Incremental regularization to compensate biased teachers in incremental learning

Learning control for complex technical systems needs a suitable trade-off between requiring little modelling efforts, fast learning and safety considerations. Incremental learning by the Directed Self-Learning strategy seems to be a good candidate for practical purposes. The learning stimuli are given incrementally by a law of adaptation acting as a teacher. But this teacher may be biased in several ways. This paper indicates that such a situation of biased teachers in incremental learning can be compensated by regularization. But in this context, regularization has to be incremental. Such an incremental regularization scheme is formally analyzed in order to extract engineering and design guidelines. The scheme is then demonstrated in a simulation setup of incremental function approximation with different biased teachers and compared to the cerebellar modelling articulation controller (CMAC).

Controlling the Learning Dynamics of Interacting Self-Adapting Systems

Complex technical systems like robots or cars are composed of many embedded subsystems to control partial dynamical effects of the whole system. In order to ease engineering and to cope with changing environmental or system properties, these subsystems need to be self-adapting. But for this to be feasible, they cannot observe the theoretically required state space of the whole system. Instead, they need to work with a reduced set of input variables. This leads to a lack of information which may induce unintended, dynamic interactions between the self-adaptation processes. Within this paper, a method is proposed in order to control the self-adaptation processes and to fight these interactions in a goal directed way. The approach is investigated on a real robotic arm.