Discovering and managing similarity knowledge in temporal case-based reasoning systems

Abstract

CBR is now a mature technology within the area of Artificial intelligence (AI). The majority of research in CBR focuses on handling time invariant attributes, whereas the temporal ones are either ignored or over simplified. The main reason for this is that current CBR processes such as similarity determination, indexing, adaptation and knowledge maintenance cannot satisfactorily handle time related data. This is a shortcoming of current CBR systems (see chapter 2) as time is an important and pervasive concept in the real world and therefore by default, highly relevant to many of the domains CBR has been applied to.

A novel, CBR, domain independent indexing technique called Discretised-Highest Similarity (DHS) is proposed and extensively evaluated in this thesis. Four modifications are presented (D-HSM,D-HSW, D-HSE & D-HSEW), which address accuracy and efficiency issues related to case-bases with different mixtures of attribute types (numeric, nominal, temporal). All modifications are developed with the vision of being integrated into the retrieval knowledge container for handling cases with non-temporal (see chapter 3) and temporal (see chapter 5) attributes under the same framework.

This study shows that the D-HS retrieval accuracy is directly related to the distribution of normalized numeric attribute values in a case-base. According to it a case-base, which has attribute values that are evenly distributed, is more likely to produce a better accuracy result with D-HSM and D-HSW, than case-bases with an uneven distribution. However, such modifications as D-HSE and D-HSEW are more likely to perform better than the other two modifications when a distribution is uneven, due to the intelligent approach for creating indexing intervals.

The intelligent selection mechanism D-HSS is proposed (see chapter 4) to analyse the distribution of normalized numeric attribute values in a case-base and depending on the obtained result to select the most suitable indexing strategy that optimises efficiency and accuracy for a case-base. It is developed with the vision of being integrated into the maintenance knowledge container to perform a choice between the most efficient and the most accurate D-HS modification depending on the distribution of numeric attributes in a case-base and overall efficiency goals of the CBR System.

In summary (see chapter 6), this thesis has concluded the use of each D-HS modification depending on objectives of developing CBR system and case-base types. It also has presented a potential direction for the further improvement of the developed approaches.

Date of Award	Jun 2010
Original language	English
Supervisor	David Patterson (Supervisor) & Hui Wang (Supervisor)