Oct. 13 2005 2:00PM

Oct. 13 2005 3:30PM

In the talk, some results about application of Bayesian statistics in industrial (in a wide sense) problems are presented.

1. On the use of Bayesian Belief Networks in modelling human and organisational causes of maritime accidents

A novel approach for integrating human and organisational factors into risk analysis will be presented. This approach has been developed and applied to a case study in the marine industry, but it can be utilised in other industrial sectors. The approach consists of a BBN model of the maritime transport system that has been developed by taking into account the different actors of the maritime transport system (i.e., ship-owner, shipyard, port, and the regulator) and their mutual influences. These influences have been modelled through a set of variables whose combinations express the relevant functions performed by each actor.

2. A Bayesian approach to critical chain and buffer management

Execution of activities in due time is a critical aspect in project management since realisation of industrial plants (or other similar projects) after the contractual time causes heavy losses to the contractor, both financially and in terms of bad reputation. Different methods have been proposed in literature to control activities during projects; in particular, the authors (jointly with Palomo and Rios Insua) considered a Bayesian dynamic linear model to forecast delivery times of items provided by subcontractors. The current paper stems from work by Goldratt on critical chains and provides a Bayesian model, sounder from a mathematical point of view than Goldratt's, to determine times for realisation of activities. In this approach, each activity has to be performed within a given time and a buffer time is introduced at the end of each "critical path" to recover from excesses in the previously scheduled activities. In particular, past data and experts' opinions are used to forecast the (posterior predictive) distribution of time realisations for each individual activity and for all the activities within a critical path. Times for each activity and the final buffer are assigned according to some quantiles of such distributions.

3. A Bayesian multi-fractal model with application to analysis and simulation of disk usage

Evaluating the performance of storage systems is a key aspect of the design and implementation of computers with heavy I/O workloads. Due to the difficulties and expenses involved in obtaining actual measurements of disk usage, disk performance simulators are usually fed with synthetic traces. As in many areas of computing and telecommunications, the time processes of disk usage exhibit dependencies that span over long ranges. In addition to the slow decaying of the auto-correlation function, the series have a bursty behaviour that can not usually be captured by commonly used times series methods. Also, when the number of packets, as opposed to inter-arrival times, is the variable of interest, a distribution with a point mass at zero has to be considered, since there is positive probability that no activity is observed during a given unit of time. Multiplicative cascade models have been considered in the literature as a way of capturing the bursty behaviour of the series. Such models have multi-fractal properties providing a rich structure that is able to capture the behaviour of series of disk usage. In this lecture we present a
Bayesian multi-scale modelling framework consisting of a multiplicative cascade, based on Haar wavelet transforms. We analyse data recorded by the Department of Storage and Content of HP Laboratories. We use the data to estimate the parameters of the model and provide predictive distributions from which to draw simulations that mimic the actual data. Also, we consider a Bayesian estimator of the multi-fractal spectrum of the original data.


Dr. Fabrizio Ruggeri
CNR IMATI
Milano
Italy
fabrizio@mi.imati.cnr.it

Pravidelný seminář Rozhodování a řízení za neurčitosti. (Contact person: Miroslav Kárný)