AI researchers in diverse areas have recently embraced the decision-theoretic framework as providing a theoretical foundation for building rational agents. At the same time there has been increasing interest in building mixed-initiative and interactive planning and decision making systems.

Decision-theoretic problem solving will ultimately require interactive model assessment in that probabilistic and value models for realistic domains will be too large to elicit in full, and the solutions proposed by such a system will typcially require additional explanation and justification.

Building an interactive decision-theoretic problem solver raises a number of problems concerning elicitation of the domain model and presentation of the results. While standard techniques are available for eliciting probability and utility models, the elicitation task is typically time consuming and tedious. Elicitation in decision analysis has required specification of a complete model, even though much of the model may be irrelevant to the problem actually being solved. Furthermore, decision-analytic elicitation requries the skill of an expert to identify what information is important and what simplifying assumptions are appropriate.

Once the model has been elicited and the appropriate analysis performed, the results must be presented to the user in an easily intelligible form and one that facilitates communicating additional requirements to the system if the user is not satisfied with the results. In complex domains that involve uncertainty, this can be highly challenging. The proposed symposium will provide a forum for identifying key problems to be addressed and techniques for solving them.

Issues to be addressed include:

1. The nature of interaction required by various applications.
2. Elicitation of probability and utility models
   • Representations that faciliate elicitation, including qualitative representations, grapical representations, and hybrid representations.
   • Strategies for eliciting information from the user. These include issues such as to how to identify important information, what simplifying assumptions might be appropriate, how to verify those assumptions, and how to elicit information incrementally.
   • Combining knowledge elicitation from the user with knowledge acquisition from databases. Issues include how to determine when interactive elicitation is needed to complement or supplement knowledge acquired from databases; how to interleave the two processes; and how to combine the acquired knowledge.
3. Inference
   • Performing inference with qualitative representations and with incomplete models. Combining qualitative and quantitative models.
   • Transformation between representations for communication with the user and representations for efficient inference.
   • Incorporation of constraints other than probability and utility into the decision-theoretic reasoning process.
4. Presentation of results to the user
   • Effective display of the results of decision-theoretic analysis.
   • Facilitation of user feedback.
   • Explanation generation.
   • Sensitivity analysis.

EVIDENCE THAT THE SYMPOSIUM IS OF INTEREST

The AAAI Spring Symposium on Decision-Theoretic Planning held in 1994 attracted approximately 75 participants. This year spring symposia were held on "Qualitative Preferences in Deliberation and Practical Reasoning" and on "Computational Models for Mixed Initiative Interaction". In 1996 a AAAI workshop was held on elicitation of probabilities for Bayesian networks. All these forums have partially addressed the issues in this proposal but none has provided the focus that we are proposing.