Full text
| |
| Author(s): |
Ricardo Barz Sovat
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2002-03-07 |
| Examining board members: |
André Carlos Ponce de Leon Ferreira de Carvalho;
Jorge Muniz Barreto;
Teresa Bernarda Ludermir;
Maria Carolina Monard;
Marcio Rillo
|
| Advisor: | André Carlos Ponce de Leon Ferreira de Carvalho; Sandra Maria Aluisio |
| Abstract | |
In this thesis, the integration of two Artificial Intelligence paradigms, Case-Based Reasoning and Artificial Neural Networks, is studied. The research is performed in two dififerent directions. First, the study of applying the Case-Based Reasoning methodology to the problem of choosing and configuring an Artificial Neural Network model. In second place, the feasibility of introducing an Artificial Neural Network inside of a case-based system working cycle. The solutions to the problem of choosing and configuring an Artificial Neural Network model have a strong empirical component. There is no available formalized knowledge that provides substance for an unified implementation process of those systems, known as connectionist. The solution quality depends upon the designer skill in adjusting a set of several related parameters. The Case-Based Reasoning methodology has its fundaments on the idea that an efficient expert is not a rule processor, but a collector of practical experiences, well succeeded or not. So, the methodology becomes very sound to be applied to domains where the knowledge is more difluse and it is difficult to make it explicit. From those observations, it is proposed the problem representation as a typical design task and it is established a strategy to apply the methodology in its solution. In the other direction, the choice of the best solution, inside the Case-Based Reasoning methodology, depends upon the existence of good processes that allow the transformation of a former solution into an adequate solution to the present problem. Those processes can take profit, as is shown along the work, of a good generalization capacity of the acquired knowledge. In most of the actual systems, those transformations, or adaptations, are accomplished by production rales. Those rules also demand a high degree of knowledge acquisition in domains not always well structured. Artificial Neural Networks have as a strong characteristic the ability of learning from examples, extract intrinsic features from datasets and to generalize this acquired knowledge. This ability gives them credentials to be good options in substituting rule based systems. What could be considered a weak characteristic of the Neural Networks, its leak of justifications to make its associations or predictions, does not constitute a barrier to its introduction in this specific point of the Case-Based Reasoning cycle. Based on those premises, this work suggests a neurosymbolic hybrid approach as a mechanism of retrieving and adapting cases inside this cycle. In order to provide a testing tool, it was also created a Case-Based Reasoning development environment. (AU) | |