The recent years have seen many developments in uncertainty reasoning taking place around Bayesian Networks
(BNs). BNs allow fast and efficient probabilistic reasoning. One of the key issues that researchers have
faced in using a BN is determining its parameters and structure for a given problem. Many techniques have
been developed for learning BN parameters from a given dataset pertaining to a particular problem. Most
of the methods developed for learning BN parameters from partially observed data have evolved around the
Expectation-Maximization (EM) algorithm. In its original form, EM algorithm is a deterministic iterative two-step
procedure that converges towards the maximum-likelihood (ML) estimates.
The EM algorithm mainly focuses on learning BN parameters from imperfect data where some of the values are
missing. However in many practical applications, partial observability results in a wider range of imperfections,
e.g., uncertainties arising from incomplete, ambiguous, probabilistic, and belief theoretic data. Moreover, while
convergence is to their ML estimates, the EM algorithm does not guarantee convergence to the underlying true
In this paper, we propose an approach that enables one to learn BN parameters from a dataset containing
a wider variety of imperfections. In addition, by introducing an early stopping criterion together with a new
initialization method to the EM-algorithm, we show how the BN parameters could be learnt so that they are
closer to the underlying true parameters than the converged ML estimated parameters.
Numerous applications of topical interest call for knowledge discovery and classification from information that may be inaccurate and/or incomplete. For example, in an airport threat classification scenario, data from heterogeneous sensors are used to extract features for classifying potential threats. This requires a training set that utilizes non-traditional information sources (e.g., domain experts) to assign a threat level to each training set instance. Sensor reliability, accuracy, noise, etc., all contribute to feature level ambiguities; conflicting opinions of experts generate class label ambiguities that may however indicate important clues. To accommodate these, a belief theoretic approach is proposed. It utilizes a data structure that facilitates belief/plausibility queries regarding “ambiguous” itemsets. An efficient <b>apriori</b>-like algorithm is then developed to extract frequent such itemsets and to generate corresponding association rules. These are then used to classify an incoming “ambiguous” data instance into a class label (which may be “hard” or “soft”). To test its performance, the proposed algorithm is compared with C4.5 for several databases from the UCI repository and a threat assessment application scenario.
Resource allocation and congestion control are two interrelated critical issues that arise in a task-oriented distributed sensor network. An effective resource management policy must account for these and their impact on the overall objectives of the network. In this paper, the viability of a virtual per-flow framework for addressing both resource allocation and congestion control in an integrated environment is demonstrated. In this framework, the resources being allocated to a physical buffer at a decision node are established by allocating and maintaining certain virtual resources to each incoming data flow. The virtual per-flow framework allows the design of controllers for each link independently of the others thus enabling a decoupled analysis and allowing one to incorporate different delay models and nonlinearities for each input data link. The effectiveness of the per-flow strategy is demonstrated via the design of a robust H<SUB>(infinity</SUB> )-norm based feedback controller that ensures extremely good tracking of a dynamically changing set-point of a decision node buffer of a distributed sensor network. The design is robust against the time-varying and uncertain nature of network-induced delays.
Research continues in exploring active control techniques to calm resonant floor vibrations. In this research, an electro- magnetic proof-mass actuator is used to deliver the control force in a single-input/multi-output control strategy. With the intent of improving the stability characteristics and the effectiveness of the SISO controller, the relative actuator mass displacement and the actuator mass velocity are added to the floor velocity output used in prior research by the author. Three separate performance indices are developed and implemented to illustrate their particular usefulness in designing an output feedback scheme for controlling pedestrian induced floor motion. The multi-output scheme has been shown analytically to further reduce steady-state acceleration amplitudes by a factor of 7 over the single-output scheme.
A digital filter which has been designed to be limit cycle free may exhibit limit cycles at the implementation stage. This is due to the inability to implement filter coefficients exactly in hardware when they are quantized to satisfy available wordlength requirements. Given a digital filter which is limit cycle free under zero input conditions, the work below presents an algorithm which finds a region in the coefficient space, about the nominal filter coefficient values, where in the filter remains limit cycle free. Furthermore the results of the algorithm will also indicate the availability of other machine representable numbers for the coefficients that fall within this robustness region. Hence one may even choose shorter wordlength registers for coefficient storage if the corresponding grid falls within the constructed robustness region.