This paper investigates the effect of learning a forward model on the performance of a statistical forward planning agent. We transform Conway's Game of Life simulation into a single-player game where the objective can be either to preserve as much life as possible or to extinguish all life as quickly as possible. In order to learn the forward model of the game, we formulate the problem in a novel way that learns the local cell transition function by creating a set of supervised training data and predicting the next state of each cell in the grid based on its current state and immediate neighbours. Using this method we are able to harvest sufficient data to learn perfect forward models by observing only a few complete state transitions, using either a look-up table, a decision tree, or a neural network. In contrast, learning the complete state transition function is a much harder task and our initial efforts to do this using deep convolutional auto-encoders were less successful.We also investigate the effects of imperfect learned models on prediction errors and game-playing performance, and show that even models with significant errors can provide good performance.

Concerning the variety of challenging situations of automated driving, a simple average fusion of all available information sources is not appropriate to obtain satisfying results. Hence, this paper presents a novel framework for the road estimation task by incorporating reliability into the multi-source fusion. First, we specify the common JDL fusion model for this task and extend it at multiple levels. Secondly, we integrate an offline-trained knowledge base for the reliability assessment represented by Bayesian Network or Random Forests. Thirdly, we propose a reliability-aware fusion of various sources at the decision level by applying Dempster-Shafer theory. As a result, our system can solve conflict situations among the sources more satisfyingly. Compared to the average fusion, experiments on real world data verify that our concept can increase the overall performance of automated driving up to 8 percentage points.

For highly available automated driving, a robust road estimation is indispensable. In order to tackle the challenges of this task, many works employ a fusion of multiple sources, e.g., visually detected lane markings, leading vehicle, digital maps, etc. However, each source has certain advantages and drawbacks depending on the operational scenarios. Hence, the assumption made by many existing approaches that the sources always are equally reliable for the fusion process is inappropriate. Therefore, this work proposes a novel concept by incorporating reliabilities into the multisource fusion so that the road estimation task can alternately select only the most reliable sources. Thereby, the reliability for each source is estimated online using classifiers trained with the sensor measurements, the past performance, and the context. Using real data recordings, experimental results show that the presented reliability-aware fusion increases the availability of automated driving up to 7 percentage points compared to the average fusion.

Community and cluster detection is a popular field of social network analysis. Most algorithms focus on static graphs or series of snapshots. In this paper we present an hierarchical algorithm, which detects communities in dynamic graphs. The method is based on the shortest paths to high-connected nodes, so called hubs. Due to local message passing, we can update the clustering results with low computational effort. The used hierarchy allows to process the community detection without setting any parameters before. After processing it provides different cluster levels based on the selected threshold. The presented algorithm is compared with the Louvain method on large-scale real-world datasets with given community structure. The detected community structure is compared to the given with NMI scores. The advantage of the algorithm is the good performance in dynamic scenarios.

Genetic programming algorithms seek to find interpretable and good solutions for problems which are difficult to solve analytically. For example, we plan to use this paradigm to develop a car accident severity prediction model for new occupant safety functions. This complex problem will suffer from the major disadvantage of genetic programming, which is its high demand for computational effort to find good solutions. A main reason for this demand is a low rate of convergence. In this paper, we introduce a new genetic operator called forking to accelerate the rate of convergence. Our idea is to interpret individuals dynamically as centers of local Gaussian distributions and allow a sampling process in these distributions when populations get too homogeneous. We demonstrate this operator by extending the Cartesian Genetic Programming algorithm and show that on our examples convergence is accelerated by over 50% on average. We finish this paper with giving hints about parameterization of the forking operator for other problems.

Recognizing and analyzing change is an important human virtue because it enables us to anticipate future scenarios and thus allows us to act pro-actively. One approach to understand change within a domain is to analyze how modelsand patterns evolve. Knowing how a model changes over time is suggesting to ask: Can we use this knowledge to learn a model in anticipation, such that it better reflects the near-future characteristics of an evolving domain? In this paper we provide an answer to this question by presenting an algorithm which predicts future decision trees based ona model of change. In particular, this algorithm encompasses a novel approach to change mining which is based on analyzing the changes of the decisions made during model learning. The proposed approach can also be applied to other types of classifiers and thus provides a basis for future research. We present our first experimental results which show that anticipated decision trees have the potential to outperform trees learned on the most recent data.

Recently a new type of data source came into the focus of knowledge discovery from temporal data: interval sequences. In contrast to event sequences, interval sequences contain labeled events with a temporal extension. However, existing algorithms for mining patterns from interval sequences proved to be far from satisfying our needs. In brief, we missed an approach that, at the same time, defines support as the number of pattern instances, allows input data that consists of more than one sequence, implements time constraints on a pattern instance, and counts multiple instances of a pattern within one interval sequence. In this paper we propose a new support definition which incorporates these properties. We also describe FSMSet, an algorithm that employs the new support definition, and demonstrate its performance on field data from the automotive business.

3015xxii

This paper presents different techniques to visualize high-dimensional fuzzy rule bases in relation to the classified data. The degree of membership to influential rules can be visualized for an entire data set. This enables the observer to detect conflicting or error-prone rules as well as misclassified feature vectors. Results are shown on a benchmark data set and on a weather data set that is used to predict flight durations on a major European airport.

Recently several papers studied resampling approaches to determine the number of clusters in prototype-based clustering. The core idea underlying these approaches is that with the right choice for the number of clusters basically the same cluster structures should be obtained from subsamples of the given data set, while a wrong choice should produce considerably varying cluster structures. In this paper we investigate whether these approaches can be transferred to fuzzy clustering. It turns out that they are applicable to fuzzy clustering as well, but that not all relative cluster evaluation measures that work for crisp clustering can also be used for fuzzy clustering.

The detection of the types of local surface form deviations is a major step in the automated quality assessment of car body parts during the manufacturing process. In previous studies we compared the performance of different soft computing techniques for this purpose. We achieved promising results with regard to classification accuracy and interpretability of rule bases, even though the dataset was rather small, high dimensional and unbalanced. In this paper we reconsider the collection of training examples and their assignment to defect types by the quality experts. We attempt to minimize the uncertainty of the quality experts' subjective and error-prone labelling in order to achieve a higher reliability of the defect detection. We show that refined and more accurate classification models can be built on the basis of a preprocessed training set that is more consistent. Using a partially supervised learning strategy we can report improvements in classification accuracy.

In this paper, a typicality-based clustering algorithm is proposed: it exploits typicality degrees defined in a prototype construction framework to identify a decomposition of the dataset into homogeneous and distinct clusters and to provide characteristic representatives of the obtained clusters, so as to summarise the initial dataset. The proposed algorithm can be applied both to vectorial and non vectorial data, such as trees for instance. Tests performed on artificial and real data illustrate the interest of the proposed approach.

Multidimensional scaling provides low-dimensional visualisation of high-dimensional feature vectors. This is a very important step in data preprocessing because it helps the user to appraise which methods to use for further data analysis. But a well known problem with conventional MDS is the quadratic need of space and time. Beside this, a transformation of MDS must be completely recomputed if additional feature vectors have to be considered. The POLARMAP algorithm, presented in this paper, learns a function, similar to NeuroScale, but with lower computational costs, that maps high-dimensional feature vectors to a 2-dimensional feature space. With the obtained function even new feature vectors can be mapped to the target space

We study an extension of fuzzy learning vector quantization that draws on ideas from the more sophisticated approaches to fuzzy clustering, enabling us to find fuzzy clusters of ellipsoidal shape and differing size with a competitive learning scheme. This approach may be seen as a kind of online fuzzy clustering, which can have advantages w.r.t. the execution time of the clustering algorithm. We demonstrate the usefulness of our approach by applying it to document collections, which are, in general, difficult to cluster due to the high number of dimensions and the special distribution characteristics of the data

In fuzzy clustering soft cluster partitions are formed based on the similarity of data points to the respective cluster prototypes. Similarity is defined in terms of simultaneous closeness regarding all attributes. In some applications the values of many attributes have been measured, but a natural clustering, if it exists, occurs within a (small) subset of attributes. The remaining dimensions can be considered irrelevant. They can obscure an existing grouping and make it harder to discover the cluster structure. In probabilistic fuzzy clustering irrelevant attributes can lead to coincidental cluster centers in the worst case. We study this effect in detail as well as the robustness of different similarity functions and their possible parameterizations against irrelevant input dimensions. Empirical evidence is given for the different properties of the membership functions

Summary form only given. Graphical models have become one of the most popular tools to structure uncertain knowledge about high dimensional domains in order to make reasoning in such domains feasible. Their most prominent representatives are Bayesian networks and Markov networks, but also relational and possibilistic networks turned out to be useful in practical applications. For all types of networks several clear, correct, and efficient propagation methods have been developed, with join tree propagation and bucket elimination being among the most widely known. In practice, however, the need also arises to support a variety of additional knowledge based operations on graphical models, where revision, updating, the fusion of networks with relational rule systems, network approximation, and learning from data samples are some of the most important ones. Furthermore, it is essential to provide software tools in order to make interactive planning, reasoning, and decision making feasible, even in complex networks of real world applications. So lots of interesting research topics in this area has to be addressed. The research to be reported about here was mainly triggered by consulting of the automobile manufacturer Daimler-Chrysler and Volkswagen Group, where graphical models are now established for several tasks. In opposite to many competitors, these two manufacturers favour a marketing policy that provides a maximum degree of freedom in choosing individual specifications of vehicles. That is, considering personal preferences, a customer may select from a large variety of options, each of which is taken from a so called item family that characterizes a certain line of equipment. Typical examples include body variants, engines, gearshifts, door layouts, seat coverings, radios, and navigation systems. In case of the VW Golf there are about 200 families with typically 4 to 8 values each, and a total range of cardinalities from 2 up to 150. The presentation refers to new theoretical and algorithmic results on decomposable models as well as some details on industrial applications.

Real life transaction data often miss some occurrences of items that are actually present. As a consequence some potentially interesting frequent patterns cannot be discovered, since with exact matching the number of supporting transactions may be smaller than the user-specified minimum. In order to allow approximate matching during the mining process, we propose an approach based on transaction editing. Our recursive algorithm relies on a step by step elimination of items from the transaction database together with a recursive processing of transaction subsets. This algorithm works without complicated data structures and allows us to find fuzzy frequent patterns easily.

A significant number of scientific and economic problems is characterised by a large number of interrelated variables. But with larger variable number, the domain under consideration may grow fast, so that analyses and reasoning become increasingly difficult. Graphical models allow to represent the combined distributions compactly and are suitable for dealing with uncertain and incomplete information. We describe their application to a problem of industrial planning. We also demonstrate how the iterative planning process can be supported by allowing the users to adapt the model using revision and updating operators. Moreover we discuss the problem of inconsistent inputs.

We explore how techniques that were developed to improve the training process of artificial neural networks can be used to speed up fuzzy clustering. The basic idea of our approach is to regard the difference between two consecutive steps of the alternating optimization scheme of fuzzy clustering as providing a gradient, which may be modified in the same way as the gradient of neural network back-propagation is modified in order to improve training. Our experimental results show that some methods actually lead to a considerable acceleration of the clustering process.

Today the method for surface quality analysis of exterior car body panels is still characterized by manual detection of local form deviations and evaluation by experts. The new approach presented in this paper is based on 3-D image processing. A major step in this process is the classification of the different kinds of surface form deviations. For this purpose, we used neuro-fuzzy classification and other soft computing techniques and compared the performance of the different approaches. Although the dataset was rather small, high-dimensional and unbalanced, we achieved promising results with regard to classification accuracies and interpretability of rule bases.

Graphical models - especially probabilistic networks like Bayes networks and Markov networks - are very popular to make reasoning in high-dimensional domains feasible. Since constructing them manually can be tedious and time consuming, a large part of recent research has been devoted to learning them from data. However, if the dataset to learn from contains imprecise information in the form of sets of alternatives instead of precise values, this learning task can pose unpleasant problems. In this paper, we survey an approach to cope with these problems, which is not based on probability theory as the more common approaches like, e.g., expectation maximization, but uses the possibility theory as the underlying calculus of a graphical model. We provide semantic foundations of possibilistic graphical models, explain the rationale of possibilistic decomposition as well as the graphical representation of decompositions of possibility distributions and finally discuss the main approaches to learn possibilistic graphical models from data.

Due to their rather intuitive and understandable application fuzzy if-then rules are a popular basis for classifiers. The use of linguistic variables eases the readability and interpretability of the rule base. In many practical applications huge amounts of data are available. However, these are often unlabeled and the user must manually assign labels. The idea of semi-supervised learning is to use as much labeled data as available and try to additionally exploit the structure in the unlabeled data. We describe an approach to enable semi-supervised learning for (neuro-) fuzzy systems.

We explore an approach to possibilistic fuzzy clustering that avoids a severe drawback of the conventional approach, namely that the objective function is truly minimized only if all cluster centers are identical. Our approach is based on the idea that this undesired property can be avoided if we introduce a mutual repulsion of the clusters, so that they are forced away from each other. We develop this approach for the possibilistic fuzzy c-means algorithm and the Gustafson-Kessel algorithm.

A major topic of recent research in graphical models has been to develop algorithms to learn them from a dataset of sample cases. However, most of these algorithms do not take into account that learned graphical models may be used for time-critical reasoning tasks and that in this case the time complexity of evidence propagation may have to be restricted, even if this can be achieved only by accepting approximations. In this paper we suggest a simulated annealing approach to learn graphical models with hypertree structure, with which the complexity of the popular join tree evidence propagation method can be controlled at learning time by restricting the size of the cliques of the learned network.

The problem of combining pieces of information issued from several sources can be encountered in various fields of application. This paper aims at presenting the different aspects of information fusion in different domains, such as databases, regulations, preferences, sensor fusion, etc., at a quite general level. We first present different types of information encountered in fusion problems, and different aims of the fusion process. Then we focus on representation issues which are relevant when discussing fusion problems. An important issue is then addressed, the handling of conflicting information. We briefly review different domains where fusion is involved, and describe how the fusion problems are stated in each domain. Since the term fusion can have different, more or less broad, meanings, we specify later some terminology with respect to related problems, that might be included in a broad meaning of fusion. Finally we briefly discuss the difficult aspects of validation and evaluation. © 2001 John Wiley & Sons, Inc.

In this paper we study methods for predicting the German stock index DAX. The idea is to use the information provided by several different information sources. We consider two different types of information sources: (1) human experts who formulate their knowledge in the form of rules, and (2) databases of objective measurable time series of financial values. It is shown how to fuse these different types of knowledge by using neuro-fuzzy methods. We present experimental results that demonstrate the usefulness of these new concepts. In the second part of the paper we present methods for the evaluation and combination of different methods for DAX prediction by using a probabilistic assessment methodology. © 2001 John Wiley & Sons, Inc.

Although probabilistic networks and fuzzy clustering may seem to be disparate areas of research, they can both be seen as generalizations of naive Bayes classifiers. If all descriptive attributes are numeric, naive Bayes classifiers often assume an axis-parallel multidimensional normal distribution for each class. Probabilistic networks remove the requirement that the distributions must be axis-parallel by taking covariances into account where this is necessary. Fuzzy clustering tries to find general or axis-parallel distributions to cluster the data. Although it neglects the classification information, it can be used to improve the result of the above mentioned methods by removing the restriction to only one distribution per classification.

Fuzzy classification rules are widely considered to be a well-suited representation of classification knowledge, as they allow readable and interpretable rule bases. This paper discusses the shapes of the resulting classification borders under consideration of different types of fuzzy sets, rule bases and t-norms, and thus which class distributions can be represented by such classification systems. We focus on discussing how antecedent pruning influences the classification behaviour of fuzzy classifiers. Our main goal is to give the potential user an insight into the classification behaviour of fuzzy classifiers. For this, 2D and 3D visualisations are mainly used to illustrate the cluster shapes and the borders between distinct classes.

The paper deals with the use of fuzzy set theory in succession planning. In the whole process of management planning, succession planning is one part which plays an important role. Selecting a candidate for a concrete position is particularly difficult if you have a lot of alternatives. The paper proposes a way of flexible querying in ordinary databases. In order to improve query results, the SQL systax is extended to allow a kind of "weak" query with the help of fuzzy methods. This sort of imprecise information retrieval is more convenient for managers who are responsible for personnel recruitment and planning. Queries simplify the decision process by pre-selecting the alternatives. The results of this work are included into a fuzzy based software tool, which is used by the VW AG to select successors for vacant positions.

The way engineers use fuzzy control in real world applications is often not coherent with an understanding of the control rules as logical statements or implications. In most cases fuzzy control can be seen as an interpolation of a partially specified control function in a vague environment, which reflects the indistinguishability of measurements or control values. In this paper the authors show that equality relations turn out to be the natural way to represent such vague environments and they develop suitable interpolation methods to obtain a control function. As a special case of our approach the authors obtain Mamdani's model and can justify the inference mechanism in this model and the use of triangular membership functions not only for the reason of simplified computations, and they can explain why typical fuzzy partitions are preferred. The authors also obtain a criterion for reasonable defuzzification strategies. The fuzzy control methodology introduced in this paper has been applied successfully in a case study of engine idle speed control for the Volkswagen Golf GTI.>