Summary. Many complex systems have characteristics which vary over time. Consider for example, the problem of modelling a river as the seasons change or adjusting the setup of a machine as it ages, to enable it to stay within predefined tolerances. In such cases offline learning limits the capability of an algorithm to accurately capture a dynamic system, since it can only base predictions on events that were encountered during the learning process. Model updating is therefore required to allow the model to change over time and to adapt to previously unseen events. In the sequel we introduce an extended version of the fuzzy Bayesian prediction algorithm [6] which learns models incorporating both uncertainty and fuzziness. This extension allows an initial model to be updated as new data becomes available. The potential of this approach will be demonstrated on a real-time flood prediction problem for the River Severn in the UK.
[1]
Jonathan Lawry,et al.
River Flow Modelling Based on Fuzzy Labels
,
2004
.
[2]
Peter C. Young,et al.
Data assimilation in the identification of flood inundation models: derivation of on-line muti-step ahead predictions in flows.
,
2004
.
[3]
Jonathan Lawry,et al.
A framework for linguistic modelling
,
2004,
Artif. Intell..
[4]
Jonathan Lawry,et al.
Linguistic Modelling Using a semi-naive Bayes Framework
,
2002
.
[5]
David D. Lewis,et al.
Naive (Bayes) at Forty: The Independence Assumption in Information Retrieval
,
1998,
ECML.
[6]
R. Jeffrey.
The Logic of Decision
,
1984
.
[7]
Jonathan Lawry.
Modelling and Reasoning with Vague Concepts
,
2006,
Studies in Computational Intelligence.