Abstract not found

A predictive modular neural network method is applied to the problem of unsupervised time series segmentation. The method consists of the concurrent application of two algorithms: one for source identification, the other for time series classification. The source identification algorithm discovers the sources generating the time series, assigns data to each source and trains one predictor for each source. The classification algorithm recursively computes a credit function for each source, based on the competition of the respective predictors, according to their predictive accuracy; the credit function is used for classification of the time series observation at each time step. The method is tested by numerical experiments.

We present a new on-line multi-model algorithm for parameter estimation of timevarying nonlinear systems. The time-variation is captured by assuming that the system parameters change according to a MarkovJan mechanism. The algorithm postulates a finite nmnber of possible values of the system parameter and computes recursively the credit function of each parameter value, according to its predictive accuracy. A convergence analysis of the algorithm is presented which indicates that the algorithm estimates correctly the parameter value, in the time intervals between source switchings. This conclusion is corroborated by numerical experiments.

In data mining, concept drift refers to the phenomenon that the underlying model (or concept) is changing over time. The aim of this paper is twofold. First, we propose a fundamental characterization and quantification of different types of concept drift. The proposed theory enables a rigorous investigation of learning system performance on streamed data. In particular, we investigate the impact of different amounts and types of concept drift on evolutionary classification systems focusing on the learning classifier system approach. We compare performance of one Pittsburgh-type system, GAssist, which learns in batch mode using windowing techniques, with a Michigan-type system, XCS, which is a natural online learner. The results show that both systems are able to handle the various concept drifts well. Behavioral differences are discussed revealing task dependencies, representation dependencies as well as dynamics dependencies. Discussions and conclusions outline the path towards more detailed measures for problem dynamics in the data mining realm. 1

In this thesis the problem of monitoring the condition of machinery is addressed using cyclostationary and higher-order statistical analysis of vibration signals. These have the potential to provide information about the machine's condition when the signal is contaminated by noise which disrupts more traditional information extraction procedures such as power spectral analysis. Features extracted from the vibration signals can be used to classify the condition. The performance of the various diagnostic features is compared over a wide range of classification systems ranging from simple fault detection thresholds to artificial neural networks. The higher-order statistics of the vibrations are exploited to generate features which are unaffected by Gaussian noise. Simple zero-lag statistical features, suitable for real-time implementation, and higher-order spectral features are evaluated for their diagnostic ability. The cyclostationarity of the vibration signals of rotating machines is i...

In this paper we present an application of predictive modular neural networks (PREMONN) to short term load forecasting. PREMONNs are a family of probabilistically motivated algorithms which can be used for time series prediction, classification and identification. PREMONNs utilize local predictors of several types (e.g. linear predictors or artificial neural networks) and produce a final prediction which is a weighted combination of the local predictions; the weights can be interpreted as Bayesian posterior probabilities and are computed online. The method is applied to short term load forecasting for the Greek Public Power Corporation dispatching center of Crete, where PREMONN outperforms conventional prediction techniques. 2 Problem Formulation We are given a sequence y t , t=1,2, ... , where (for each t) y t has dimensions 24 1; each of the y t components corresponds to the load of a particular hour of the day on day no. t. The predictors have the general form y t =f(y t-1 , y t-2...

In this paper we explore some aspects of the problem of on-line, unsupervised learning of a switching time series, i.e. a time series which is generated by a combination of several, alternately activated sources. This learning problem can be solved by a two-stage approach: (a) separating of the incoming data to several datasets (one dataset corresponding to each source); (b) developing one model per dataset (i.e. one model per source). We introduce a general data allocation methodology which combines the two steps into an iterative scheme: existing models compete for the incoming data; data assigned to each model are used to refine the model. We distinguish between two modes of data allocation: in parallel data allocation every incoming datablock is allocated to the model with lowest prediction error; in serial data allocation the incoming datablock is allocated to the first model with prediction error below a prespecified threshold. We present sufficient conditions for asymptotically correct allocation of the data. We also present numerical experiments to support our theoretical analysis. 1