Thesis (Ph.D.)--Utrecht University, 2006. / Includes bibliographical references (p. [133]-137).

<p>Initially an analytical closed-form inverse kinematics solution for a 5 DOF robotic arm was developed and implemented. This analytical solution proved not to meet the accuracy required for the shape sorting puzzle setup used in the COSPAL (COgnitiveSystems using Perception-Action Learning) project [2]. The correctness of the analytic model could be confirmed through a simulated ideal robot and the source of the problem was deemed to be nonlinearities introduced by weak servos unable to compensate for the effect of gravity. Instead of developing a new analytical model that took the effect of gravity into account, which would be erroneous when the characteristics of the robotic arm changed, e.g. when picking up a heavy object, a learning approach was selected.</p><p>As learning method Locally Weighted Projection Regression (LWPR) [27] is used. It is an incremental supervised learning method and it is considered a state-ofthe-art method for function approximation in high dimensional spaces. LWPR is further combined with visual servoing. This allows for an improvement in accuracy by the use of visual feedback and the problems introduced by the weak servos can be solved. By combining the trained LWPR model with visual servoing, a high level of accuracy is reached, which is sufficient for the shape sorting puzzle setup used in COSPAL.</p>

<p>A technique to detect deviations in sets of systems in a self-organized way is described in this work. System features are extracted to allow compact representation of the system. Distances between systems are calculated by computing distances between the features. The distances are then stored in an affinity matrix. Deviating systems are detected by assuming a statistical model for the affinities. The key idea is to extract features and and identify deviating systems in a self-organized way, using nonlinear techniques for the feature extraction. The results are compared with those achieved with linear techniques, (principal component analysis).</p><p>The features are computed with principal curves and an isometric feature mapping. In the case of principal curves the feature is the curve itself. In the case of isometric feature mapping is the feature a set of curves in the embedding space. The similarity measure between two representations is either the Hausdorff distance, or the Frechet distance. The deviation detection is performed by computing the probability of each system to be observed given all the other systems. To perform reliable inference the Bootstrapping technique was used.</p><p>The technique is demonstrated on simulated and on-road vehicle cooling system data. The results show the applicability and comparison with linear techniques.</p>

Coordination is essential to achieving good performance in cooperative multiagent systems. To date, most work has focused on either implicit or explicit coordination mechanisms, while relatively little work has focused on the benefits of combining these two approaches. In this work we demonstrate that combining explicit and implicit mechanisms can significantly improve coordination and system performance over either approach individually. First, we use difference evaluations (which aim to compute an agent's contribution to the team) and stigmergy to promote implicit coordination. Second, we introduce an explicit coordination mechanism dubbed Intended Destination Enhanced Artificial State (IDEAS), where an agent incorporates other agents' intended destinations directly into its state. The IDEAS approach does not require any formal negotiation between agents, and is based on passive information sharing. Finally, we combine these two approaches on a variant of a team-based multi-robot exploration domain, and show that agents using a both explicit and implicit coordination outperform other learning agents up to 25%. / Graduation date: 2012

Bayesian networks are used for building intelligent agents that act under uncertainty. They are a compact representation of agents' probabilistic knowledge. A Bayesian network can be viewed as representing a factorization of a full joint probability distribution into the multiplication of a set of conditional probability distributions. Independence of causal influence enables one to further factorize the conditional probability distributions into a combination of even smaller factors. The efficiency of inference in Bayesian networks depends on how these factors are combined. Finding an optimal combination is NP-hard. We propose a new method for efficient inference in large Bayesian networks, which is a combination of new representations and new combination algorithms. We present new, purely multiplicative representations of independence of causal influence models. They are easy to use because any standard inference algorithm can work with them. Also, they allow for exploiting independence of causal influence fully because they do not impose any constraints on combination ordering. We develop combination algorithms that work with heuristics. Heuristics are generated automatically by using machine learning techniques. Empirical studies, based on the CPCS network for medical diagnosis, show that this method is more efficient and allows for inference in larger networks than existing methods. / Graduation date: 1999

The study of physical activity is important in improving people���s health as it can help people understand the relationship between physical activity and health. Accelerometers, due to its small size, low cost, convenience and its ability to provide objective information about the frequency, intensity, and duration of physical activity, has become the method of choice in measuring physical activity. Machine learning algorithms based on the featurized representation of accelerometer data have become the most widely used approaches in physical activity prediction. To improve the classification accuracy, this thesis first explored the impact of the choice of data (raw vs processed) as well as the choice of features on the performance of various classifiers. The empirical results showed that the machine learning algorithms with strong regularization capabilities always performed better if provided with the most comprehensive feature set extracted from raw accelerometer signal. Based on the hypothesis that for some time series, the most discriminative information could be found at subwindows of various sizes, the Subwindow Ensemble Model (SWEM) was proposed. The SWEM was designed for the accelerometer-based physical activity data, and classified the time series based on the features extracted from subwindows. It was evaluated on six time series datasets. Three of them were accelerometer-based physical activity data, which the SWEM was designed for, and the rest were different types of time series data chosen from other domains. The empirical results indicated a strong advantage of the SWEM over baseline models on the accelerometerbased physical activity data. Further analysis confirmed the hypothesis that the most discriminative features could be extracted from subwindows of different sizes, and they were effectively used by the SWEM. / Graduation date: 2013

This paper describes the initial results of a project to create a self-supervised algorithm for learning object segmentation from video data. Developmental psychology and computational experience have demonstrated that the motion segmentation of objects is a simpler, more primitive process than the detection of object boundaries by static image cues. Therefore, motion information provides a plausible supervision signal for learning the static boundary detection task and for evaluating performance on a test set. A video camera and previously developed background subtraction algorithms can automatically produce a large database of motion-segmented images for minimal cost. The purpose of this work is to use the information in such a database to learn how to detect the object boundaries in novel images using static information, such as color, texture, and shape. / Singapore-MIT Alliance (SMA)

Co-training is a semi-supervised learning method that is designed to take advantage of the redundancy that is present when the object to be identified has multiple descriptions. Co-training is known to work well when the multiple descriptions are conditional independent given the class of the object. The presence of multiple descriptions of objects in the form of text, images, audio and video in multimedia applications appears to provide redundancy in the form that may be suitable for co-training. In this paper, we investigate the suitability of utilizing text and image data from the Web for co-training. We perform measurements to find indications of conditional independence in the texts and images obtained from the Web. Our measurements suggest that conditional independence is likely to be present in the data. Our experiments, within a relevance feedback framework to test whether a method that exploits the conditional independence outperforms methods that do not, also indicate that better performance can indeed be obtained by designing algorithms that exploit this form of the redundancy when it is present. / Singapore-MIT Alliance (SMA)

This thesis presents Pathway Informed Analysis (PIA), a classification method for predicting disease states (diagnosis) from metabolic profile measurements that incorporates biological knowledge in the form of metabolic pathways. A metabolic pathway describes a set of chemical reactions that perform a specific biological function. A significant amount of biological knowledge produced by efforts to identify and understand these pathways is formalized in readily accessible databases such as the Kyoto Encyclopedia of Genes and Genomes. PIA uses metabolic pathways to identify relationships among the metabolite concentrations that are measured by a metabolic profile. Specifically, PIA assumes that the class-conditional metabolite concentrations (diseased vs. healthy, respectively) follow multivariate normal distributions. It further assumes that conditional independence statements about these distributions derived from the pathways relate the concentrations of the metabolites to each other. The two assumptions allow for a natural representation of the class-conditional distributions using a type of probabilistic graphical model called a Gaussian Markov Random Field. PIA efficiently estimates the parameters defining these distributions from example patients to produce a classifier. It classifies an undiagnosed patient by evaluating both models to determine the most probable class given their metabolic profile. We apply PIA to a data set of cancer patients to diagnose those with a muscle wasting disease called cachexia. Standard machine learning algorithms such as Naive Bayes, Tree-augmented Naive Bayes, Support Vector Machines and C4.5 are used to evaluate the performance of PIA. The overall classification accuracy of PIA is better than these algorithms on this data set but the difference is not statistically significant. We also apply PIA to several other classification tasks. Some involve predicting various manipulations of the metabolic processes performed in experiments with worms. Other tasks are to classify pigs according to properties of their dietary intake. The accuracy of PIA at these tasks is not significantly better than the standard algorithms.

In a fully automated manufacturing environment, instant detection of condition state of the cutting tool is essential to the improvement of productivity and cost effectiveness. In this paper, a tool condition monitoring system (TCM) via machine learning (ML) and machine ensemble (ME) approach was developed to investigate the effectiveness of multisensor fusion when machining 4340 steel with multi-layer coated and multi-flute carbide end mill cutter. Feature- and decision-level information fusion models utilizing assorted combinations of sensors were studied against selected ML algorithms and their majority vote ensemble to classify gradual and transient tool abnormalities. The criterion for selecting the best model does not only depend on classification accuracy but also on the simplicity of the implemented system where the number of features and sensors is kept to a minimum to enhance the efficiency of the online acquisition system. In this study, 135 different features were extracted from sensory signals of force, vibration, acoustic emission and spindle power in the time and frequency domain by using data acquisition and signal processing modules. Then, these features along with machining parameters were evaluated for significance by using different feature reduction techniques. Specifically, two feature extraction methods were investigated: independent component analysis (ICA), and principal component analysis (PCA) and two feature selection methods were studied, chi square and correlation-based feature selection (CFS). For various multi-sensor fusion models, an optimal feature subset is computed. Finally, ML algorithms using support vector machine (SVM), multilayer perceptron neural networks (MLP), radial basis function neural network (RBF) and their majority voting ensemble were studied for selected features to classify not only flank wear but also breakage and chipping. In this research, it has been found that utilizing the multisensor feature fusion technique under majority vote ensemble gives the highest classification performance. In addition, SVM outperformed other ML algorithms while CFS feature selection method surpassed other reduction techniques in improving classification performance and producing optimal feature sets for different models.