Claudia Krull

A new approach for touch gesture recognition: Conversive Hidden non-Markovian Models

Abstract With the current boom of touch devices the recognition of touch gestures is becoming an important field of research. Performing such gestures can be seen as a stochastic process, as there can be many little differences between different executions. Therefore stochastic models like Hidden Markov Models have already been applied to gesture recognition. Although the modelling possibilities of Hidden Markov Models are limited, they achieve an acceptable recognition quality. But they have never been tested with gestures that only differ in execution speed. We propose the use of Conversive Hidden non-Markovian Models for touch gesture recognition. This extension of Hidden Markov Models enhances the modelling possibilities and adds timing features. In this paper, two touch gesture recognition systems were developed and implemented based on these two model types. Experiments with a set of similar gestures show that the proposed model class is a good and competitive alternative to Hidden Markov Models.

Virtual Stochastic Sensors: How to Gain Insight into Partially Observable Discrete Stochastic Systems

This paper introduces the idea of a Virtual Stochastic Sensor. This paradigm enables the analysis of unobservable processes in discrete stochastic systems. Just like a virtual sensor, we use physical sensor readings to deduce the value of the quantity of interest. However, both the physical sensor readings and their relationship with the quantity of interest are stochastic. Therefore the measurement of our virtual stochastic sensor is a statistical estimate of the true value. We describe a method to compute the result of the virtual stochastic sensor and show its validity and real-time capability for two example models. We also give system properties that must apply in order for the feasibility of virtual stochastic sensors, such as the sensitivity of the physical sensor output to changes in the quantity of interest. The future potential of virtual stochastic sensors is their variability. They can be used to gain insight into hidden processes of partially observable systems, using readily available data. They enable online monitoring of production lines using already recorded data to ensure optimal control and maximum production efficiency.

Virtual Stochastic Sensors for Reconstructing Job Shop Production Workflows

Virtual stochastic sensors (VSS) can reconstruct the system behavior of partially observable systems that may contain concurrent non-Markovian activities. These systems can for example occur in industrial production. In this paper, we apply VSS to reconstructing workflows in a job shop. The example application was designed in cooperation with logistics experts to resemble a real job shop. In order to deal with probabilistic decisions in the workflows, we extended Hidden non-Markovian Models to include immediate transitions and modified the corresponding Proxel-based behavior reconstruction algorithm accordingly. Experimental data was acquired using a setup including a printed layout, RFID sensors and a central data collection unit. We were able to reconstruct the actual workflows from the acquired sensor data, which for the first time shows an application of VSS to real measurement data. This first practical application, and the extension of the modeling paradigm takes us forward to our goal of a realistically applicable method for behavior reconstruction based on partial system observations.

Using hidden non-Markovian Models to reconstruct system behavior in partially-observable systems

Many complex technical systems today have some basic protocol capability, which is used for example to monitor the quality of production output or to keep track of oil pressure in a modern car. The recorded protocols are usually used to detect deviations from some predefined standards and issue warnings. However, the information in such a protocol is not sufficient to determine the source or cause of the problem, since only part of the system is being observed. In this paper we present an approach to reconstruct missing information in only partially-observable stochastic systems based only on recorded system output. The approach uses Hidden non-Markovian Models to model the partially-observable system and Proxel-based simulation to analyze the recorded system output. Experiments were conducted using a production line example. The result of the analysis is a set of possible system behaviors that could have caused the recorded protocol, including their probabilities. We will show that our approach is able to reconstruct the relevant information to determine the source of non-standard system behavior. The combination of Hidden non-Markovian Models and Proxel-based simulation holds the potential to reconstruct unobserved information from partial or even noisy output protocols of a system. It adds value to the information already recorded in many production systems today and opens new possibilities in the analysis of inherently only partially-observable systems.

Interpretation and optimization of material flow via system behavior reconstruction

Abstract In this article, a method for a scalable autonomous data acquisition for an analysis and optimization of production systems based on interpretation of the material flow within small and medium-sized manufacturing enterprises is presented. The data is acquired locally and combined centrally to interpret the material flow as a basis for the optimization of the material flow as well as individual processes. When it is not completely observable for efficiency reasons or due to technical restrictions, one can also reconstruct relevant but unobservable system behavior based on system knowledge and actual measurements. A validation of the method is carried out in a company maintaining engines. The application of the model shows that with the presented method it is possible to reduce buffers in the production, optimize transportation routes and reduce waiting and therefore cycle times in job shop productions for an increasing productivity.

Evaluating a New Conversive Hidden non-Markovian Model Approach for Online Movement Trajectory Verification.

This paper presents further research on an implemented classification and verification system that employs a novel approach for stochastically modelling movement trajectories. The models are based on Conversive Hidden non-Markovian Models that are especially suited to mimic temporal dynamics of time series as in contrast to the relative Hidden Markov Models(HMM) and the dynamic time warping(DTW) method, timestamp information of data are an integral part. The system is able to create trajectory models from examples and is tested on signatures, doodles and pseudo-signatures for its verification performance. By using publicly available databases comparisons are made to evaluate the potential of the system. The results reveal that the system already performs similar to a general DTW approach on doodles and pseudo-signatures but does not reach the performance of specialized HMM systems for signatures. But further possibilities to improve the results are discussed.

Reconstructing model parameters in partially-observable discrete stochastic systems

The analysis of partially-observable discrete stochastic systems reconstructs the unobserved behavior of real-world systems. An example for such a system is a production facility where indistinguishable items are produced by two machines in stochastically distributed time intervals and are then tested by a single quality tester. Here, the source of each defective item can be reconstructed later based solely on the time-stamped test protocol. While existing algorithms can reconstruct various characteristics of the unobserved behavior, a fully specified discrete stochastic model needs to exist. So far, model parameters themselves cannot be reconstructed. In this paper, we present two new approaches that enable the reconstruction of some unknown parameter values in the model specification, namely constant probabilities. Both approaches are shown to work correctly and with acceptable computational effort. They are a first step towards general model parameter inference for partially-observable discrete stochastic systems.

A Markov model for multi-criteria multi-person decision making

This paper describes a new algorithm for the evaluation of alternatives by a group of decision makers according to multiple criteria. The algorithm is motivated by the need to quickly evaluate a large number of ideas in the early stages of an innovation process, when little or no information about the ideas is available. The algorithm is based on a Markov chain model which is derived from pairwise comparisons of ideas. The steady-state solution of this Markov chain yields a ranking vector for the alternatives. The algorithm is similar to the ldquoPageRankrdquo method used by Google. The new algorithm does not require absolute values and allows assignment of weights both to the decision makers and to the evaluation criteria.

A Conversive Hidden Non-Markovian Model Approach for 2D and 3D Online Movement Trajectory Verification

A novel approach for stochastically modelling movement trajectories is presented that has already been implemented and evaluated for classification scenarios in previous research and in this article its applicability to verification scenarios is analysed. The models are based on Conversive Hidden non-Markovian Models that are especially suited to mimic temporal dynamics of time series. In contrast to the popular Hidden Markov Models (HMM) and the dynamic time warping (DTW) method, timestamp information of the data is an integral part. A verification system is presented that create trajectory models from several examples and its verification performance is deduced from experiments on different data sets including signatures, doodles, pseudo-signatures and hand gestures recorded with a Kinect. The results are compared to other publications and they reveal that the developed system already performs similar to a general DTW approach, but expectedly does not yet reach the quality of specialized HMM systems. It is also shown that the system can be applied to three dimensional data and further possibilities to improve the results are discussed.

Change Detection of Model Transitions in Proxel Based Simulation of CHnMMs

To analyze discrete stochastic models, Virtual Stochastic Sensors were developed at the Otto-von-Guericke-University Magdeburg. This procedure makes it possible to reconstruct the behavior of a broader class of hidden models, like Conversive Hidden non-Markovian Models, in a very efficient way. One assumption of this approach is that the distribution functions, which describe the state changes of the system, are time-homogeneous. However, this assumption is not always true when it comes to real world problems.