IRIDIA[ Main | Projects | People | Meeting | Library ] IRIDIA


Implementation of Uncertain Reasoning


Intro

In order to make belief function theory applicable to real problems, we have developed efficient algorithms based on the Valuation- based framework.


Valuation-based systems

Efficient computer implementations of uncertain reasoning can be obtained by using the so-called network-based approaches like the valuation-based system (VBS) of Shenoy and Shafer. In VBS, knowledge is represented by a network of nodes representing entities of the domain of discourse and their states, and links representing relations between these entities. These relations can express logical constraints or quantified uncertainty. For dealing with a problem, we consider all the relevant elements and build the corresponding nodes and links in the network. Then we associate values called valuations with these links for representing our knowledge. In VBS, we can represent uncertain knowledge using different uncertainty formalisms, including probability theory, belief function theory, possibility theory, etc..

The inference mechanism of a VBS is based on two operators called combination and marginalization. By the combination of all the valuations in a network, we get the joint valuation on the whole domain of interest. Marginalization is then used to focus this joint valuation on the domain of interest, and compute the uncertainty valuation restricted to that domain. When there are many nodes in a VBS, brute force computation of the joint valuation is computationally intractable. However, it is often possible to compute the marginals on the domain of interest using local computation, therefore strongly reducing the computational complexity. Local computation can be used when: (1) the problem can be decomposed, i.e., it can be described as a combination of sub-problems; and (2) the operators of combination and marginalization satisfy certain axioms - these axioms are satisfied by classical uncertainty representations such as probability, belief function, and possibility theories.


TresBel

In order for belief function theory to be applicable to real problems within the VBS framework, we have developed efficient algorithms for implementing local computation scheme. These algorithms are used in TresBel, our implementation of the transferable belief model (TBM). TresBel is implemented in a modular way, in Lisp, and is enriched with a graphical interface. TresBel can accept both joint and conditional belief functions. We use TresBel as an open and practical tool to experiment different algorithms, and to verify their efficiency and adequateness to model problems of uncertain reasoning. Besides, we have developed two strategies for the explanations of origin of the computed beliefs. One is based on the sensitivity analysis, the other is based on the measure of information contents of a belief function.

The initial VBS was adapted to joint beliefs. In order to cope with conditional beliefs, we have also developed a network that accepts conditional belief functions. It is more natural and much easier for the users to provide conditional beliefs for representing the relations between the nodes. Besides, computation is faster and memory requirements are smaller when using conditional beliefs instead of joint beliefs.


PULCINELLA

We have then developed a second tool based on the VBS formalism, called PULCINELLA. PULCINELLA generalizes TresBel to cover other uncertainty formalisms, including possibility theory, probability theory, and classical propositional logic. PULCINELLA has been used to comparatively test the adequacy of these formalisms in modelling real-world problems. See the PULCINELLA Home Page for more details.


ISDAT

VBS can be also used for representing and solving Bayesian decision problems. Problems are represented in an extended VBS framework, and the solution still benefits from the local computation. Generalizing PULCINELLA, we have implemented a system, called VBSD, for Bayesian decision analysis in VBS. The TBM provides a mental two-level structure for representing uncertainty, one that characterises the state of belief of the agent and where beliefs are represented by belief functions, the other that characterises the behavior of the agent and where a decision is to be made using the classical expected utility theory. In our decision support system based on the TBM, we have integrated the programs TresBel and VBSD into ISDAT, an integrated architecture that allows decision making under uncertainty.


[ Hong XU, Alessandro SAFFIOTTI, Philippe SMETS ]


Selected references

Saffiotti A. and Umkehrer E.
Pulcinella: A General Tool for Propagating Uncertainty in Valuation Networks
in D'Ambrosio B. D., Smets Ph. And Bonissone P. P. eds. Proc. 7th Uncertainty in Artificial Intelligence, pp. 323-331, San Mateo, Ca.: Morgan Kaufmann 1991.
Xu H.
An Efficient Tool for Reasoning with Belief Functions
in Bouchon-Meunier B.,Valverde L. and Yager R. R. eds. Uncertainty in Intelligent Systems pp. 215-224, North-Holland, Amsterdam 1993.
Xu H. and Kennes R.
Steps towards an Efficient Implementation of Dempster-Shafer Theory,
in Yager R. R., Fedrizzi M., and Kacprzyk J. eds. Advances in the Dempster-Shafer Theory of Evidence, pp. 153-174, New York: Wiley 1994.
Xu H. and Smets Ph.
Reasoning in Evidential Networks with Conditional Belief Functions International Journal of Approximate Reasoning, 1995.
Xu H., Hsia Y-T. and Smets Ph.
Transferable Belief Model for Decision Making in Valuation-Based Systems, IEEE Transactions on Systems, Man andCybernetics, 1996.


IRIDIA[ Main | Projects | People | Meeting | Library ] IRIDIA

Last updated Jan 15, 1996 | Comments to webmaster