In an experimental environment, we simulated the situation of a user who gives speech input to a system while walking through an airport. The time pressure on the subjects and the requirement to navigate while speaking were manipulated orthogonally. Each of the 32 subjects generated 80 utterances, which were coded semi-automatically with respect to a wide range of features, such as filled pauses. The experiment yielded new results concerning the effects of time pressure and cognitive load on speech. To see whether a system can automatically identify these conditions on the basis of speech input, we had this task performed for each subject by a Bayesian network that had been learned on the basis of the experimental data for the other subjects. The results shed light on the conditions that determine the accuracy of such recognition. 1 Background and Issues This paper is an experimental follow-up to the UM99 paper by Berthold and Jameson ([2]). Those authors argued the follo...

One feature of intelligent user interfaces is an ability to make decisions that take into account a variety of factors, some of which may depend on the current situation. This article focuses on one general approach to such decision making: Predict the consequences of possible system actions on the basis of prior empirical learning, and evaluate the possible actions, taking into account situation-dependent priorities and the tradeoffs between the consequences. This decisiontheoretic approach is illustrated in detail with reference to an example decision problem, for which models for decision making were learned from experimental data. It is shown how influence diagrams and methods of decision-theoretic planning can be applied to arrive at empirically well-founded decisions. This paradigm is then compared with two other paradigms that are often employed in intelligent user interfaces. Finally, various possible ways of learning (or otherwise deriving) suitable decision-theoretic models are dis- cussed.

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...

This paper explores semi-qualitative probabilistic networks (SQPNs) that combine numeric and qualitative information. We first show that exact inferences with SQPNs are NP PP-Complete. We then show that existing qualitative relations in SQPNs (plus probabilistic logic and imprecise assessments) can be dealt effectively through multilinear programming. We then discuss learning: we consider a maximum likelihood method that generates point estimates given a SQPN and empirical data, and we describe a Bayesian-minded method that employs the Imprecise Dirichlet Model to generate set-valued estimates. 1

ABSTRACT: User-adaptive systems are interactive software systems that spontaneously adapt to their individual users– for example, to their interests or their work habits. First, we briefly characterize this type of system and consider its relationship to the broader category of smart adaptive systems. We then give an overview of the computational techniques that are used to realize user-adaptation, ranging from data-based machine learning techniques to theory-based decisiontheoretic models. Special attention is given to the question of possible synergies: What methods that have been developed for user-adaptive systems might profitably be transferred to other smart adaptive systems? And what techniques from the latter field deserve increased attention in connection with user-adaptivity?

Bayesian networks have gained increasing attention in recent years. One key issue in Bayesian networks (BNs) is parameter learning. When training data is incomplete or sparse or when multiple hidden nodes exist, learning parameters in Bayesian networks (BNs) becomes extremely difficult. Under these circumstances, the learning algorithms are required to operate in a high-dimensional search space and they could easily get trapped among copious local maxima. This paper presents a learning algorithm to incorporate domain knowledge into the learning to regularize the otherwise ill-posed problem, to limit the search space, and to avoid local optima. Unlike the conventional approaches that typically exploit the quantitative domain knowledge such as prior probability distribution, our method systematically incorporates qualitative constraints on some of the parameters into the learning process. Specifically, the problem is formulated as a constrained optimization problem, where an objective function is defined as a combination of the likelihood function and penalty functions constructed from the qualitative domain knowledge. Then, a gradient-descent procedure is systematically integrated with the E-step and M-step of the EM algorithm, to estimate the parameters iterativelyuntil it converges. The experiments with both synthetic data and real data for facial action recognition show 2 our algorithm improves the accuracy of the learned BN parameters significantly over the conventional EM algorithm. I.

As molecular biology continues to evolve in the direction of high-throughput collection of data, it has become increasingly necessary to develop computational methods for analyzing observed data that are at once both sophisticated enough to capture essential features of biological phenomena and at the same time approachable in terms of their application. We demonstrate how graphical models, and Bayesian networks in particular, can be used to model genetic regulatory networks. These methods are well-suited to this problem owing to their ability to model more than pair-wise relationships between variables, their ability to guard against over-fitting, and their robustness in the face of noisy data. Moreover, Bayesian network models can be scored in a principled manner in the presence of both genomic expression and location data. We develop methods for extending Bayesian network semantics to include edge annotations that allow us to model statistical dependencies between biological factors with greater refinement. We derive principled methods for scoring these annotated Bayesian networks. Using these models in the presence...

Abstract. One of the central questions addressed in the project READY was that of how a system can automatically recognize situationally determined resource limitations of its user—in particular, time pressure and cognitive load. This chapter summarizes most of the work done in READY on this topic, presenting as well some previously unpublished results. We first consider why on-line recognition or resource limitations can be useful by discussing the ways in which a system might adapt its behavior to perceived resource limitations. We then summarize a number of approaches to the recognition problem that have been taken in READY and other projects, before focusing on one particular approach: the analysis of features of a user’s speech. In each of two similarly structured experiments, we created four experimental conditions that varied in terms of whether the user was (a) required to produce spoken utterances quickly or not; and (b) navigating within a simulated airport terminal or standing still. In the second experiment, additional distraction was caused by continuous loudspeaker announcements. The speech produced by the experimental subjects (32 in each experiment) was coded in terms of 7 variables. We report on the extent to which each of these variables was influenced by the subjects ’ resource limitations. We also trained dynamic Bayesian networks on the resulting data in order to see how well the information in the users ’ speech could serve as evidence as to which condition the user had been in. The results yield information about the accuracy that can be attained in this way and about the diagnostic value of some specific features of speech. 1

In this paper, we derive a method to refine a Bayes network diagnostic model by exploiting constraints implied by expert decisions on test ordering. At each step, the expert executes an evidence gathering test, which suggests the test’s relative diagnostic value. We demonstrate that consistency with an expert’s test selection leads to non-convex constraints on the model parameters. We incorporate these constraints by augmenting the network with nodes that represent the constraint likelihoods. Gibbs sampling, stochastic hill climbing and greedy search algorithms are proposed to find a MAP estimate that takes into account test ordering constraints and any data available. We demonstrate our approach on diagnostic sessions from a manufacturing scenario. 1