The WinMine Toolkit

 

Max Chickering (Microsoft Research)

 

The WinMine Toolkit is set of tools that allow you to build Bayesian networks and dependency networks from data. These tools, which are for academic use, can be downloaded for free off of the web. For my presentation, I will be available to show those interested how to use the tools.

 

 

 

GeNIe and SMILE: a development environment for building decision-theoretic models

 

Denver Dash and Changhe Yuan

 

SMILE (Structural Modeling, Inference, and Learning Engine) is a fully portable library of C++ classes implementing graphical decision-theoretic methods, such as  Bayesian networks and influence diagrams, directly amenable to inclusion in intelligent systems.  Its Windows user interface, GeNIe is a versatile and user-friendly development environment for graphical decision-theoretic models.  Both modules, developed at the Decision Systems Laboratory, University of Pittsburgh, have been made available to the community in July 1998 (the download site is http://www.sis.pitt.edu/~genie).

 

 

 

B-Course, a web-based data analysis tool

 

Petri Myllymaki, Petri Kontkanen, Hannes Wettig (Complex Systems Computation Research Group, Helsinki Institute for Information Technology)

 

B-Course is a web-based tool for Bayesian and causal modeling. It can be used as a data analysis tool for any research where dependence, classification or causal models based on sample data are of interest. B-Course can also be used as an interactive tutorial, which provides you with data sets that have been prepared in advance. B-Course can be freely used for educational and research purposes at address http://b-course.hiit.fi.

 

 

 

A class of location-based decisioning tasks that is now feasible via an MDP approach

 

Mark E. Plutowski (Cyber Technologies Laboratory, Sony Electronic Inc.)

 

I evaluated four MDP solvers upon a class of stochastic planning tasks drawn from a location-based application scenario. The problems involve generating the optimal visitation sequence across a number of geographically located "stores" at which one or more valuable resources are located with some probability, governed by a utility function based on travel times and acquisition rewards. Two of these solvers use a factored MDP approach that mitigates the state space explosion problem by lumping states together by value. The algorithm underlying these two techniques is based upon a dynamic Bayesian network formulation, utilizing the algebraic decision diagram (ADD) as the primary data structure. A third solver utilizes a symbolic heuristic approach that utilizes the same data structures, but which obtains greater efficiencies by ignoring unreachable states during the forward expansion phase of value iteration. A fourth solver utilizes a full reachability forward expansion phase starting from a given set of states. I settled upon the symbolic heuristic approach as the main MDP solver, using the other three for computing the heuristic. The approach exploits the strong reachability structure of the task in the heuristic as well as the value function computation. I conclude that it is now feasible to attack moderately sized task instances from this class of problems. I discuss some practical lessons learned along the way in applying these algorithms and identify some new challenges.

 

 

 

REES: Reasoning Engine(s) Evaluation Shell

 

Radu Marinescu, Kalev Kask and Rina Dechter (University of California, Irvine)

 

We introduce a software environment to support research and development in the area of both deterministic and non-deterministic reasoning. This environment –REES (Reasoning Engine(s) Evaluation Shell) has a plug-in oriented architecture that promotes reuse of existing software components and allows for the comparison and evaluation of alternative technologies. The third release of the REES system and technical documentation is now available at http://www.ics.uci.edu/~radum/rees.html.

 

 

 

PRISM: A first order modeling system

 

Taisuke Sato (Tokyo Institute of Technology / CREST JST)

Neng-Fa Zhou (The City University of New York)

 

PRISM1.6 is a symbolic-statistical modeling language for modeling complex phenomena.  It is an extension of logic programming language B-Prolog that offers the user (1) probabilistic built-ins for constructing logical-probabilistic models in terms of logic programs and (2) a fast EM learner, called graphical EM, that learns parameters embedded in programs from observed facts.  We are applying PRISM1.6 to several applications including statistical natural language processing (HPSG), modeling Shogi (Japanese chess) players, and linkage analysis.

 

 

 

Software for simulation and discovery of causality, and investigation of causal hypotheses

 

Richard Scheines, Joe Ramsey, Matt Easterday (Carnegie Mellon University)

 

Tetrad 4 supports the construction and discovery of Bayes Nets, Structural Equation Models, and Time Series Models (in particular genetic regulatory networks with microarray measurements). Algorithms include the PC algorithm, the FCI algorithm, the CCD algorithm (linear cyclic models), and a genetic algorithm for linear DAGs. Causality Lab is software for simulating the scientific process of investigating causal hypotheses. Based on a stored but hidden Bayes Net, this software allows the user to set-up and carry out passive observational studies and randomized trials, judge the independence relations that hold in such studies, hypothesize causal models that would explain such evidence, and compare predictions against the evidence.

http://www.phil.cmu.edu/projects/tetrad_download/

http://www.phil.cmu.edu/projects/causality-lab/

 

 

 

VIBES: A Variational Inference Engine for Bayesian Networks

 

John Winn (Cambridge University)

Christopher M. Bishop (Microsoft Research, Cambridge)

 

In recent years variational methods have become a popular tool for approximate inference and learning in a wide variety of probabilistic models. For each new application, however, it is currently necessary first to derive the variational update equations, and then to implement them in application-specific code. Each of these steps is both time consuming and error prone. In this presentation we demonstrate a general-purpose inference engine called VIBES (`Variational Inference for Bayesian Networks'), which allows a wide variety of probabilistic models to be implemented and solved variationally without recourse to coding. New models are specified via a graphical interface analogous to a drawing package. VIBES then automatically generates and solves the variational equations. We illustrate the power and flexibility of VIBES using examples from Bayesian mixture modelling.