Scope
The Bayesian Advisor Project is concerned with automating decision making and constraint checking aspects
of advising. The project is composed of several members working together on various aspects of
the problem. For instance, there is a tool to build the initial Bayesian Probability Network
that will model the domain, a database system that stores various information about the bayesian
network, and a tool that ultimately conducts the planning with the network. Our part of the project
deals with modifying and implementing the Probability Elicitation Tool. The PET is
an application that allows experts in various domains (the running example is in the advisor
implementation) to associate probablities in the bayesian network that will be used in planning.
The initial framework is already in place but will have to be updated and modified greatly before
it will be working at the standards it is expected to.
Objectives
There are various items that must be implemented to fullfill the goals of the Probability
Elicitation Tool:
-
The PET must be modified to correctly communicate with the current Semi-Structured
Probabilistic Database Management System.
-
The PET needs to implement consistency checking both at the current input phase
in the domain and over the entire scope of the domain.
-
The PET also needs to allow for elicitation of specific types of probability elicitation
functions (most important of which is noisy-or).
-
The PET interface needs to be modified for better usability.
-
If time allows, the PET needs to be able to elicit combination methods based
on Mahoney's and Lanskey's research on Bayes net fragments.
Project Specifications
User Interface
A picture of the original Probability Elication Tool can be seen
here. Pictures of the future PET will be
posted at a later date.
Problem Decomposition
There are several issues that need to be taken into consideration dealing with the implementation
of the various objects in the probability elicitation tool:
-
There is a huge body of psychological research that deals with how best to check for
consistency within data elicited from experts. Sometimes the discrepancies are so large that
a fusion of the data must be done to ensure the best possible accuracy.
-
The database the PET is connected to has already been implemented and is maintained by
another Bayesian Advisor Project member (Wenzhong). There are several XML schemas, special
purpose APIs, and protocols that must be thoroughly understood to ensure proper communication.
-
There are various probability distribution functions that experts can use to provide a
guide line for comparing their data against. Some of these include the bell curve, geometric curves,
min/max, and noisy-or. All of these distributions will have to be researched and understood
before we can implement them into the PET.
-
A main issue that occurs in elicitation tools is the inconsistencies that arise from experts
due to boredom from the whole elicitation process. To overcome this problem the PET needs to
have its interface redesigned to allow for easier use and greater visual appeal.
-
The extra functionality that may be implemented into the PET is the use of combination
methods to break a Bayesian network into pieces. This allows the expert to pull out those
pieces of the network that are currently of interest prior to performing computation on the data.
Data Dictionary
The Data Dictionary contains the API's for all of the
Java code behind PET. It can be found in the downloads section.
All original work on PET including APIs and
documentation were written Jiangyu Li. Updates by Williams,
Cornett, and Wong are noted where applicable.
Downloads and Documentation
Data Dictionary
Presentation Slides: .ppt
.pdf
Overview Documentation: .doc
.pdf
|
