Resumen de A dynamic adaptive framework for improving case-based reasoning system performance
Fernando Orduña Cabrera
An optimal performance of a Case-Based Reasoning (CBR) system means, the CBR system must be efficient both in time and in size, and must be optimally competent. The efficiency in time is closely related to an efficient and optimal retrieval process over the Case Base of the CBR system. Efficiency in size means that the Case Library (CL) size should be minimal. Therefore, the efficiency in size is closely related to optimal case learning policies, optimal meta-case learning policies, optimal case forgetting policies, etc. On the other hand, the optimal competence of a CBR system means that the number of problems that the CBR system can satisfactorily solve must be maximum. To improve or optimize all three dimensions in a CBR system at the same time is a difficult challenge because they are interrelated, and it becomes even more difficult when the CBR system is applied to a dynamic or continuous domain (data stream). In this thesis, a Dynamic Adaptive Case Library framework (DACL) is proposed to improve the CBR system performance coping especially with reducing the retrieval time, increasing the CBR system competence, and maintaining and adapting the CL to be efficient in size, especially in continuous domains. DACL learns cases and organizes them into dynamic cluster structures. The DACL is able to adapt itself to a dynamic environment, where new clusters, meta-cases or prototype of cases, and associated indexing structures (discriminant trees, k-d trees, etc.) can be formed, updated, or even removed. DACL offers a possible solution to the management of the large amount of data generated in an unsupervised continuous domain (data stream). In addition, we propose the use of a Multiple Case Library (MCL), which is a static version of a DACL, with the same structure but being defined statically to be used in supervised domains. The thesis work proposes some techniques for improving the indexation and the retrieval task. The most important indexing method is the NIAR k-d tree algorithm, which improves the retrieval time and competence, compared against the baseline approach (a flat CL) and against the well-known techniques based on using standard k-d tree strategies. The proposed Partial Matching Exploration (PME) technique explores a hierarchical case library with a tree indexing-structure aiming at not losing the most similar cases to a query case. This technique allows not only exploring the best matching path, but also several alternative partial matching paths to be explored. The results show an improvement in competence and time of retrieving of similar cases. Through the experimentation tests done, with a set of well-known benchmark supervised databases. The dynamic building of prototypes in DACL has been tested in an unsupervised domain (environmental domain) where the air pollution is evaluated. The core task of building prototypes in a DACL is the implementation of a stochastic method for the learning of new cases and management of prototypes. Finally, the whole dynamic framework, integrating all the main proposed approaches of the research work, has been tested in simulated unsupervised domains with several well-known databases in an incremental way, as data streams are processed in real life. The conclusions outlined that from the experimental results, it can be stated that the dynamic adaptive framework proposed (DACL/MCL), jointly with the contributed indexing strategies and exploration techniques, and with the proposed stochastic case learning policies, and meta-case learning policies, improves the performance of standard CBR systems both in supervised domains (MCL) and in unsupervised continuous domains (DACL).
