Listas KBM2L para la síntesis de conocimiento en sistemas de ayuda a la decisión

Abstract

The implementation, evaluation and exploitation of Decision Support Systems by means of Bayesian Networks and Influence Diagrams, among other reasoning models, imply the use of tables with diversified information. Among them we focus on the conditional probability tables that represent the probabilistic relationships among variables and the tables of the optimal decisions resulting from the model evaluation. The tables, that can be very complex, include structured knowledge from the application domains over a set of variables of the probabilistic graphical model. Under the name of KBM2L we introduce a technique to build the knowledge base of the decision support system. We try to exploit the KBM2L list as a useful tool for the knowledge representation of the system that includes the model graph, the utility and probability models and the evaluation output. The graphical representation is qualitative and intuitive and then the users can easily access the knowledge if they are experts on the problem. On the other hand, the quantitative models of probability and utility and the evaluation output do not show easily the knowledge because it is coded numerically, in the case of these models, and due to the huge size of the optimal decision tables, in the case of the evaluation output. Both aspects do not allow us to recognize the main variables and relationships that describe the knowledge and explain the results. While the tables can be regarded as static objects or entities, KBM2L lists are dynamic knowledge representations. A specific list configuration determines the ability of knowledge explanation, the eficiency to solve queries to the decision support system from many di®erents points of view and the memory complexity required to manage the knowledge base. The structure of the list allows us to reveal the granularity of knowledge from tables while the configurations show us the role of the model variables in the inferred evaluation. The granularity provides procedures to structure and understand better the knowledge that the system hosts in its tables. The role of the variables in the di®erent contexts and in the whole model give us a mechanism to generate explanations of knowledge and of the system proposals and also the sensitivity analysis of the model itself. After the introduction we show the foundations of the KBM2L list for knowledge representation and describe the problems of the optimal representation search and several proposals of solution. We face a combinatorial optimization problem that is dealt with algorithms and methods adapted to our objective in the framework of metaheuristics. Next, we show the application of these techniques to optimal decision tables and conditional probability tables of the influence diagram. Finally, we propose to perform a model sensitivity analysis by means of the natural extension of the usual KBM2L list with the meaningful parameters.