The problem of making optimal decisions in uncertain conditions is central to Artificial Intelligence. If the state of the world is known at all times, the world can be modeled as a Markov Decision Process (MDP). MDPs have been studied extensively and many methods are known for determining optimal courses of action, or policies. The more realistic case where state information is only partially observable, Partially Observable Markov Decision Processes (POMDPs), have received much less attention. The best exact algorithms for these problems can be very inefficient in both space and time. We introduce Smooth Partially Observable Value Approximation (SPOVA), a new approximation method that can quickly yield good approximations which can improve over time. This method can be combined with reinforcement learning methods, a combination that was very effective in our test cases. 1 Introduction Markov Decision Processes (MDPs) have proven to be useful abstractions for a variety of problems. W...

The Self-Organizing Map (SOM) is an efficient tool for visualization of multidimensional numerical data. In this paper, an overview and categorization of both old and new methods for the visualization of SOM is presented. The purpose is to give an idea of what kind of information can be acquired from different presentations and how the SOM can best be utilized in exploratory data visualization. Most of the presented methods can also be applied in the more general case of first making a vector quantization (e.g. k-means) and then a vector projection (e.g. Sammon's mapping).

Model-based clustering techniques have been widely used and have shown promising results in many applications involving complex data. This paper presents a unified framework for probabilistic model-based clustering based on a bipartite graph view of data and models that highlights the commonalities and differences among existing model-based clustering algorithms. In this view, clusters are represented as probabilistic models in a model space that is conceptually separate from the data space. For partitional clustering, the view is conceptually similar to the ExpectationMaximization (EM) algorithm. For hierarchical clustering, the graph-based view helps to visualize critical/important distinctions between similarity-based approaches and model-based approaches.

We report on the implementation of two neural network algorithms for visuomotor control of an industrial robot (Puma 562). The first algorithm uses a vector quantization technique, the "neural-gas" network, together with an error correction scheme based on a Widrow-Hoff-type learning rule. The second algorithm employs an extended self-organizing feature map algorithm. Based on visual information provided by two cameras, the robot learns to position its end effector without an external teacher. Within only 3000 training steps, the robot-camera system is capable of reducing the positioning error of the robot's end effector to approximately 0.1 percent of the linear dimension of the work space. By employing adaptive feedback the robot succeeds in compensating not only slow calibration drifts, but also sudden changes in its geometry. Hardware aspects of the robot-camera system are discussed.

In this paper we test the Self-Organizing Map (SOM) on the problem of predicting chaotic time-series (speci cally Mackey-Glass series) with local linear models de ned separately for each of the prototype vectors of the SOM. We see that the method achieves good results. This together with the capabilities of the SOM make itavaluable tool in exploratory data mining. 1

Abstract not found

Spatiotemporal connectionist networks (STCN's) comprise an important class of neural models that can deal with patterns distributed both in time and space. In this paper, we widen the application domain of the taxonomy for supervised STCN's recently proposed by Kremer (2001) to the unsupervised case. This is possible through a reinterpretation of the state vector as a vector of latent (hidden) variables, as proposed by Meinicke (2000). The goal of this generalized taxonomy is then to provide a nonlinear generative framework for describing unsupervised spatiotemporal networks, making it easier to compare and contrast their representational and operational characteristics. Computational properties, representational issues and learning are also discussed and a number of references to the relevant source publications are provided. It is argued that the proposed approach is simple and more powerful than the previous attempts, from a descriptive and predictive viewpoint. We also discuss the relation of this taxonomy with automata theory and state space modeling, and suggest directions for further work.

We propose a hierarchical self-organizing neural network ("HiGS") with adaptive architecture and simple topological organization. This network combines features of Fritzke's Growing Cell Structures and traditional hierarchical clustering algorithms. The height and width of the tree structure depend on the user-specified level of error desired, and the weights in upper layers of the network do not change in later phases of the learning algorithm. Parameters such as node deletion rate are adaptively modified by the learning algorithm. 1. Introduction Connectionist learning systems often face the stability-plasticity dilemma [2]. Most unsupervised neural network learning algorithms are stable with respect to their topology and plastic with respect to weight vector adaptations; such is the case in Kohonen's topology-preserving self-organizing map (SOM) [4]. An exception is Fritzke's Growing Cell Structures (GCS) network [1], which is much more plastic in that nodes may be inserted ...

We present a rapidly learning neural control architecture for sensory-based navigation of a mobile robot and compare the learning dynamics and the navigation behavior in the context of different implemented network approaches and learning schemes. Our control architecture is a combination of i) alternative vector quantization techniques (Neural gas and Kohonen feature map) for optimal clustering and categorizing of continuous input data spaces and ii) a neural implementation of the Q-learning, a very efficient reinforcement learning method for the choice of the appropriate actions. Our simulation experiments in an artificial environment of changeable geometrical complexity demonstrate that a robot, utilizing this control scheme, can learn the desired behavior rapidly, irrespective of the chosen contradictory navigation tasks. Moreover, we can show that only simultaneous learning schemes develop a kind of `functional categorizing' of sensory situations. Only they are capable of acquiring knowledge about the sensorial consequences of executed actions from the beginning.

this paper we concentrate on the power and stability of several popular clustering algorithms under the condition that the correct number of clusters is known. Artificial data sets modeled to mimic typical situations from tourism marketing are constructed in Section 2. The structure of these data sets is described in several scenarios, and artificial binary data are generated accordingly. These data, ranging from very simple to more complex, real-data-like structures, enable us to systematically analyze the "behavior" of the cluster methods. Section 3 gives an overview of all cluster methods under investigation. Section 4 describe our experimental results, comparing first all scenarios and then all cluster methods. To accomplish this task, several evaluation criteria for cluster methods are proposed. Finally, Sections 5 and 6 draw some conclusions and give an outlook on future research. 2 Generating Artificial Data from Scenarios