Decision theory has become widely accepted in the AI community as a useful framework for planning and decision making. Applying the framework typically requires elicitation of some form of probability and utility information. While much work in AI has focused on providing representations and tools for elicitation of probabilities, relatively little work has addressed the elicitation of utility models. This imbalance is not particularly justified considering that probability models are relatively stable across problem instances, while utility models may be different for each instance. Spending large amounts of time on elicitation can be undesirable for interactive systems used in low-stakes decision making and in time-critical decision making. In this paper we investigate the issues of reasoning with incomplete utility models. We identify patterns of problem instances where plans can be proved to be suboptimal if the (unknown) utility function satisfies certain conditions. We present an...

The number of knowledge-based systems that build on Bayesian belief networks is increasing. The construction of such a network however requires a large number of probabilities in numerical form. This is often considered a major obstacle, one of the reasons being that experts are reluctant to provide numerical probabilities. The use of verbal probability expressions as an additional method of eliciting probabilistic information may to some extent remove this obstacle. In this paper, we review studies that address the communication of probabilities in words and/or numbers. We then describe our own experiments concerning the development of a probability scale that contains words as well as numbers. This scale appears to be an aid for researchers and domain experts during the elicitation phase of building a belief network and might help users understand the output of the network.

Probabilistic networks are now fairly well established as practical representations of knowledge for reasoning under uncertainty, as demonstrated by an increasing number of successful applications in such domains as (medical) diagnosis and prognosis, planning, vision,

The hardest task in knowledge engineering for probabilistic graphical models, such as Bayesian networks and influence diagrams, is obtaining their numerical parameters. Models based on acyclic directed graphs and composed of discrete variables, currently most common in practice, require for every variable a number of parameters that is exponential in the number of its parents in the graph, which makes elicitation from experts or learning from databases a daunting task. In this paper, we review the so called canonical models, whose main advantage is that they require much fewer parameters. We propose a general framework for them, based on three categories: deterministic models, ICI models, and simple canonical models. ICI models rely on the concept of independence of causal influence and can be subdivided into noisy and leaky. We then analyze the most common families of canonical models (the OR/MAX, the AND/MIN, and the noisy XOR), generalizing them and offering criteria for applying them in practice. We also briefly review temporal

An underlying problem of current interface agent research is the failure to adequately address effective and efficient knowledge representations and associated methodologies suitable for modeling the users' interactions with the system. These user models lack the representational complexity to manage the uncertainty and dynamics involved in predicting user intent and modeling user behavior. A utility theory-based approach is presented for effective user intent prediction by incorporating the ability to explicitly model users' goals, the uncertainty in the users' intent in pursuing these goals, and the dynamics of users' behavior. We present an interface agent architecture, CIaA, that incorporates our approach and discuss the integration of CIaA with three disparate domains --- a probabilistic expert system shell, a natural language input database query system, and a virtual space plane ---that are being used as test beds for our interface agent research.

In designing a Bayesian network for an actual problem, developers need to bridge the gap between the mathematical abstractions o#ered by the Bayesian-network formalism and the features of the problem to be modelled. Qualitative probabilistic networks (QPNs) have been put forward as qualitative analogues to Bayesian networks, and allow modelling interactions in terms of qualitative signs. They thus have the advantage that developers can abstract from the numerical detail, and therefore the gap may not be as wide as for their quantitative counterparts. A notion that has been suggested in the literature to facilitate Bayesian-network development is causal independence. It allows exploiting compact representations of probabilistic interactions among variables in a network. In the paper, we deploy both causal independence and QPNs in developing and analysing a collection of qualitative, causal interaction patterns, called QC patterns. These are endowed with a fixed qualitative semantics, and are intended to o#er developers a highlevel starting point when developing Bayesian networks.

There are essentially two kinds of approaches for learning the structure of Bayesian Networks (BNs) from data. The first approach tries to find a graph which satis es all the constraints implied by the empirical conditional independencies measured in the data [PV91, SGS00a, Shi00]. The second approach searches through the space of models (either DAGs or PDAGs), and uses some scoring metric (typically Bayesian or some approximation, such as BIC/MDL) to evaluate the models [CH92, Hec95, Hec98, Kra98], typically returning the highest scoring model found. Our main interest is in learning BN structure from gene expression data [FLNP00, HGJY01, MM99, SGS00b]. In domains such as this, where the ratio of the number of observations to the number of variables is low (i.e., when we have sparse data), selecting a threshold for the conditional independence (CI) tests can be tricky, and repeated use of such tests can lead to inconsistencies [DD99]. Bayesian s...

Sensitivity analysis is a method to investigate the effects of varying a model's parameters on its predictions. It was recently suggested as a suitable means to facilitate quantifying the joint probability distribution of a Bayesian belief network. This article presents practical experience with performing sensitivity analyses on a belief network in the field of medical prognosis and treatment planning. Three network quantifications with different levels of informedness were constructed. Two poorly-informed quantifications were improved by replacing the most influential parameters with the corresponding parameter estimates from the well-informed network quantification; these influential parameters were found by performing one-way sensitivity analyses. Subsequently, the results of the replacements were investigated by comparing network predictions. It was found that it may be sufficient to gather a limited number of highly-informed network parameters to obtain a satisfying network quant...

The discipline of enterprise architecture advocates the use of models to support decision-making on enterprise-wide information system issues. In order to provide such support, enterprise architecture models should be amenable to analyses of various properties, as e.g. the availability, performance, interoperability, modifiability, and information security of the modeled enterprise information systems. This paper presents a software tool for such analyses. The tool guides the user in the generation of enterprise architecture models and subjects these models to analyses resulting in quantitative measures of the chosen quality attribute. The paper describes and exemplifies both the architecture and the usage of the tool. 1.

. Reasoning within such domains as engineering, science, management, or medicine is traditionally based on formal methods employing probabilistic treatment of uncertainty. It seems natural to base artificial reasoning systems in these domains on the normative foundations of probability theory. Two usual objections to this approach are (1) probabilistic inference is computationally intractable in the worst case, and (2) probability theory is incomprehensible for humans and, hence, probabilistic systems may be hardly usable. The first objection has been addressed effectively in the last decade by a variety of efficient exact and approximate schemes for probabilistic reasoning, applied in several practical systems. In this paper, I review the state of the art with respect to the second objection. First I argue that the observed discrepancies between human and probabilistic reasoning and the anticipated difficulties in building user interfaces are not a good reason for rejecting probabilit...