| Many problems are related to decision making. Because of the complexity and uncertainty of decision-making, and confusions of people's thoughts, the researches about effective description and solving methods for decision-making under uncertain setting have been paid a great deal of attention on and become an important field in uncertain artificial intelligence.Influences diagrams (IDs) are one of the most commonly used graphical decision models for reasoning under uncertainties. They represent beliefs and preferences of decision makers about the sequences of decision-making. An ID is a formal description of the problem that can be treated by computers and a representation can be easily understood by people in all walks of life and degrees of technical proficiency. An ID thus may be regarded as a bridge between qualitative description and quantitative specification. To a decision maker, IDs are a natural representation for capturing the semantics of decision-making process with a minimum of clutters and confusions. Thus, the decision maker may use IDs to model and evaluate complex decision-making process via a compact graphical representation. Nowadays, IDs have been extensively used in many fields. However, sometimes existing influence diagrams can not meet the demands in the real complex and diversified decision-making settings. Therefore, it is necessary to extend existing influence diagrams. This thesis mainly aims to explore how to extend the influence diagrams, how to get an influence diagram and in what ways to use influence diagrams. The researches are helpful to enhance the ability of influence diagrams for representing interactive and uncertain decision-making processes, to enrich and develop the theory and technology about influence diagrams, to expand application fields of influence diagrams.The main contributions and novelties of this study can be summarized as follows:(1) To extend influence diagrams for representing interactive and uncertain decision-making processes. Existing IDs only model decision-making process for a singal decision maker, and the parameters in any ID are precise single-valued parameters. When there are many decision makers are involved in the same competition, and these decision makers'decisions may interact or affect each other's outcomes at different stages of their decision-making process or the decision- making may involve in fuzzy and uncertain events, the existing IDs can not meet demands. This study carried out two extensions: 1) incorporated static strategic game theory into influence diagrams and presented a new influence diagram, called influence diagrams with game elements (GIDs). GIDs include an additional factor, that of the choices of strategies made by other decision makers, to the list of determinants that influence the decision-making process of each decision maker, so decision makers could make more rational decisions. As the result of integration of influence diagrams and game theory, GIDs benefit from the simplicity and efficiency of influence diagrams for modeling complex decision problems and from the suitability and rationality of applying game theory for making decisions in interactive dynamic scenarios;2) used interval-valued parameters in influence diagrams and presented another new influence diagrams, called influence diagrams with interval-valued parameters (IIDs). IIDs represent decision makers'imprecision beliefs and preferences with interval-valued parameters, therefore, decision makers can use IIDs to model decision-making processes in the situation that parameters of variables are imprecise, and it is easier for them to specify interval-valued parameters than to specify precise single-valued parameters.(2) To put forward the learning method for structures of GIDs and IIDs, and for ordinal utilities of IDs. Learning IDs means automatically getting IDs by analysing data; this consists of learning structures and learning parameters. 1) This thesis put forward a method of learning structures of GIDs and ones of the IIDs by using Bayesian networks structure learning algorithms; 2) This thesis also proposed a game model for solving multiple attribute decision-making without weight information, and proposed a new approach of learning ordinal utilities IDs. This new approach avoids difficulties attached to the specification of utilities. The experimental results showed that the new approach is helpful for the learning of IDs. (3) To propose forward evaluating algorithms for GIDs and IIDs——To Apply GIDs and IIDs to finding optimal policy. Forward evaluating an ID means to find an optimal policy which could maximize the global expected utility accumulated by local expected utility at each of decision nodes. Local maximal expected utility at each decision node does not mean maximizing global expected utility. How to find the optimal policy is a complicated combination optimization issue. Besides this, the game among decision makers and the calculation about interval conditional probability should be considered in the process for evaluating GIDs and IIDs. A genetic algorithms-based method to evaluate GIDs has been developed in this thesis, which has the advantages such as searching globally and simplicity in operation. The thesis also proposed an evaluation algorithm to generate optimal policy which may maximize the interval-valued expected utility; the proposed evaluation algorithm is based on the transformation of IIDs into Bayesian networks with interval probability parameters and on making inference in this secondary structure. The experimental studies have been performed to demonstrate and validate the approaches developed in this thesis. The experimental results showed that:1) in the interact decision setting, GIDs are helpful for making more rational decisions; 2) IIDs are the expansion of IDs with precise single-valued parameters.(4) To propose backward evaluating algorithms for IDs——To Apply IDs to infering rival's private information. The rival's decision-making model was described by IDs in this thesis. According to the criterion that a rival always hope to gain the global maximum expected utility rather than gain the local maximum expected utility, and by combining decision computation and Bayesian inference, the rival's private information about cost, bankroll, tactic, and utility might be inferred through the observation on rival's decision behaviors and using GAs. The experimental results indicated that proposed method is feasible and rational.
|
PDF Full Text Request