The paper describes aHUGIN, a tool for creating adaptive systems. aHUGIN is an extension of the HUGIN shell, and is based on the methods reported by Spiegelhalter and Lauritzen (1990a). The adaptive systems resulting from aHUGIN are able to adjust the C011ditional probabilities in the model. A short analysis of the adaptation task is given and the features of aHUGIN are described. Finally a session with experiments is reported and the results are discussed.

The trajectory of a robot is monitored in a restricted dynamic environment using light beam sensor data. We have a Dynamic Belief Network (DBN), based on a discrete model of the domain, which provides discrete monitoring analogous to conventional quantitative filter techniques. Sensor observations are added to the basic DBN in the form of specific evidence. However, sensor data is often partially or totally incorrect. We show how the basic DBN, which infers only an impossible combination of evidence, may be modified to handle specific types of incorrect data which may occur in the domain. We then present an extension to the DBN, the addition of an invalidating node, which models the status of the sensor as working or defective. This node provides a qualitative explanation of inconsistent data: it is caused by a defective sensor. The connection of successive instances of the invalidating node models the status of a sensor over time, allowing the DBN to handle both persistent and intermittent faults.

Bayesian networks are directed acyclic graphs representing independence relationships among a set of random variables. A random variable can be regarded as a set of exhaustive and mutually exclusive propositions. We argue that there are several drawbacks resulting from the propositional nature and acyclic structure of Bayesian networks. To remedy these shortcomings, we propose a probabilistic network where nodes represent unary predicates and which may contain directed cycles. The proposed representation allows us to represent domain knowledge in a single static network even though we cannot determine the instantiations of the predicates before hand. The ability to deal with cycles also enables us to handle cyclic causal tendencies and to recognize recursive plans.

Automated decision making is often complicated by the complexity of the knowledge involved. Much of this complexity arises from the context sensitive variations of the underlying phenomena. We propose a framework for representing descriptive, context-sensitive knowledge. Our approach attempts to integrate categorical and uncertain knowledge in a network formalism. This paper outlines the basic representation constructs, examines their expressiveness and efficiency, and discusses the potential applications of the framework.

The leading contender is Levi's When the set contains only one function, convex conditionalization and E-admissibility reduce to their strict Bayesian counterparts. Thus, with respect to decision making and representing and updating uncertainty, convex Bay· esianism includes strict Bayesianism as a special case. There are natural constraints on probability judg-- ments that cannot be represented by convex sets of classical probability functions. Working with the convex hull of a nonconvex set of probability func-- tions may result in unnecessary indecisiveness. This is not a convex set. Judgments of irrelevance (conditional irrelevance), that is, probabilistic independence (conditional independence}, are often made, are natural to make, can be made reliably, and provide well-known computational advantages [Pearl, 1988].

To date, most probabilistic reasoning systems have relied on a fixed belief network constructed at design time. The network is used by an application program as a representation of (in)dependencies in the domain. Probabilistic inference algorithms operate over the network to answer queries. Recognizing the inflexibility of fixed models has led researchers to develop automated network construction procedures that use an expressive knowledge base to generate a network that can answer a query. Although more flexible than fixed model approaches, these construction procedures separate construction and evaluation into distinct phases. In this paper we develop an approach to combining incremental construction and evaluation of a partial probability model. The combined method holds promise for improved methods for control of model construction based on a trade-off between fidelity of results and cost of construction.

Influence diagram is a graphical representation of belief networks with uncertainty. This article studies the structural properties of a probabilistic model in an influence diagram. In particular, structural controllability theorems and structural observability theorems are developed and algorithms are formulated. Controllability and observability are fundamental concepts in dynamic systems (Luenberger 1979). Controllability corresponds to the ability to control a system while observability analyzes the inferability of its variables. Both properties can be determined by the ranks of the system matrices. Structural controllability and observability, on the other hand, analyze the property of a system with its structure only, without the specific knowledge of the values of its elements (tin 1974, Shields and Pearson 1976). The structural analysis explores the connection between the structure of a model and the functional dependence among its elements. It is useful in comprehending problem and formulating solution by challenging the underlying intuitions and detecting inconsistency in a model. This type of qualitative reasoning can sometimes provide insight even when there is insufficient numerical information in a model.

Our understanding of non linear dynamical systems and networks has made tremendous progress during the past decades. In most cases the autonomous dynamics is studied. The situation where the network is strongly driven by an external signal has so far been less investigated even though it arises in many different contexts in the natural and artificial world. Examples include networks of interacting chemicals (proteins, RNA) in a cell driven by unpredictable external chemical signals; networks of neurons driven by an external sensory input; artificial neural networks and their applications in machine learning; the response of population dynamics and ecological networks to changes in external conditions such as the weather; the responses of stock prices to economically significant news such as a company earnings, or unemployment numbers. In all these cases taking into account the external input is essential if one wants to understand correctly the dynamics, both because the external input is often large (it cannot be treated as a small perturbation), and because in some cases the systems itself has been selected according to its response to the fluctuating and unpredictable external variables. The aim of the present work is to show, through the study of a specific but important example, how mean field techniques can provide a detailed understanding of dynamical networks strongly driven by an external signal. In the mean field approach the average feedback of the variables on themselves is taken into account through a self consistent equation, while the correlations between individual variables are neglected. The apparently extremely complicated dynamics of the network is thus reduced to much simpler evolution equations for a few collective variables. Previous applications of the mean field approach to dynamical systems (but without including an external input), and in particular neural networks, include e.g.

We study the problem of learning Markov decision processes with finite state and action spaces when the transition probability distributions and loss functions are chosen adversarially and are allowed to change with time. We introduce an algorithm whose regret with respect to any policy in a comparison class grows as the square root of the number of rounds of the game, provided the transition probabilities satisfy a uniform mixing condition. Our approach is efficient as long as the comparison class is polynomial and we can compute expectations over sample paths for each policy. Designing an efficient algorithm with small regret for the general case remains an open problem.

Mean-field variational methods are widely used for approximate posterior inference in many probabilistic models. In a typical application, mean-field methods approximately compute the posterior with a coordinate-ascent optimization algorithm. When the model is conditionally conjugate, the coordinate updates are easily derived and in closed form. However, many models of interest---like the correlated topic model and Bayesian logistic regression---are nonconjuate. In these models, mean-field methods cannot be directly applied and practitioners have had to develop variational algorithms on a case-by-case basis. In this paper, we develop two generic methods for nonconjugate models, Laplace variational inference and delta method variational inference. Our methods have several advantages: they allow for easily derived variational algorithms with a wide class of nonconjugate models; they extend and unify some of the existing algorithms that have been derived for specific models; and they work well on real-world datasets. We studied our methods on the correlated topic model, Bayesian logistic regression, and hierarchical Bayesian logistic regression.