Planning for success: The interdisciplinary approach to building Bayesian models
International Journal of Approximate Reasoning
| Hi-index | 0.00 |
This thesis addresses elicitation of models for decision making under uncertainty, and the quantification of qualitative models. The thesis describes a framework for building decision support systems that are based on stochastic planning under uncertainty with constraints1. Decision support systems are interactive computer-based systems that aid decision makers in solving decision and planning problems. Decision support systems are valuable where the amount of relevant information that needs to be considered is prohibitive for the intuition of an unaided human decision maker. This dissertation describes the model building process and the support for probabilistic information management, and assumes the existence of stochastic planner, constraint solver, plan display, and plan evaluation components of the decision support system. Although there has been great success in the field of Bayesian reasoning, there has been a major challenge in eliciting probabilistic networks from experts. The primary reason for this is that many experts do not understand the probabilistic network representation of a Bayesian network. We describe a process by which anthropologists, computer scientists, and social welfare case managers collaborated to build a stochastic model of welfare advising in Kentucky, and in the process of collaboration, rethought the Bayesian network model of Markov decision processes and a designed a new knowledge elicitation pattern. We call these models Markov decision processes with actions that have results, and represent each activity in the model by a bowtie action fragment. Based on the new knowledge elicitation pattern, we have designed and implemented a high level elicitor tool. We will also look at different human computer interaction principles that need to be followed while designing such a qualitative elicitation tool. Although the high level elicitor helps in eliciting the bowtie models from the social welfare case managers, the bowties are qualitative Bayesian networks and do not have conditional probability tables to support stochastic planning. We describe a novel quantification approach that converts the qualitative bowties into factored Markov decision processes. Our quantification approach uses a canonical combination method, NOISY-MAJORITY , in order to generate conditional probability tables for the central bowtie action result node. Factored Markov decision processes have conditional probability tables. In the planning process, the planner does not need to work with all conditional probability tables at the same time. Based on the situation, the planner would select appropriate conditional probability tables from the available collection. There is a need for special approaches to management of probabilistic data, since probabilities stored in the databases must be manipulated (combined, marginalized, conditionalized) in accordance with the laws of probability theory and secondly, probabilities are often associated with objects that have more complex structure than relational tables. In this thesis, I describe how the semistructured probabilistic database management system proposed by Dekhtyar et al. is currently being used for efficient storage and retrieval of probabilistic information. Keywords. Knowledge elicitation, quantification, Bayesian network, qualitative probabilistic network, probabilistic databases. 1This project is partially supported by the National Science Foundation under Grant No. ITR-0325063.