Adapting Case-Based Reasoning for Processing Natural Phenomena Data

Abstract

The objective of this thesis is to adapt the Case-Based Reasoning methodology for processing natural phenomena data, expressed by means of time series. This adaptation mainly involves the proposal of a robust mechanism for retrieving characteristic patterns from a data collection, which are crucial for this methodology. Despite the existence of many algorithms in this field, most of them fail while processing distorted input. Unfortunately, such distortion is natural for many types of data collections, especially for measurements of natural variables such as precipitations, river discharge volume etc. The proposed approach for retrieving characteristic patterns utilizes the Voting Experts algorithm for splitting the input, the Dynamic Time Warping for dealing with distorted inaccuracies and hierarchical clustering for grouping similar sequences. Although the proposal is laid out in an abstract way, so that it could be used for different domains, this thesis demonstrates its use on the rainfall-runoff model data.