5811267-5 Elements of Uncertain Reasoning (4cu)
Course at the Department of Computer Science, University of Helsinki
Instructor: Henry Tirri, Complex Systems Computation Group
Course Projects
Mini-project II.1 "Probabilities and inference" (deadline 1.10.97 12.00)
Deliverable: a report/memo; maximum 2 pages.
Type: Individual
Credit: 5% of the grade
Describe a setting where you can use probabilities to perform inference. Choose a simple instance of that setting, i.e., give all the necessary numerical information, and give a worked-out example of this inference. For this exercise get acquainted with an introductory text in probability and statistical inference, e.g.,
Wright, Daniel B. Understanding statistics. Sage, London, 1997
Milton J. S., Arnold Jesse C. Probability and statistics in the engineering and computing sciences. McGraw-Hill, New York (NY), 1987
Trivedi, Kishor Shridharbhai. Probability and statistics with reliability, queuing, and computer science applications. Prentice-Hall, Englewood Cliff (NJ), 1982.
Hinkle Dennis E., Wiersma William, Jurs Stephen G. Applied statistics for the behavioral sciences. Houghton Mifflin, Boston (MA), 1988
or any other introductory text book on the topic.
Project II.2 "Prediction of Bernoulli trials" (deadline 20.10.97 12.00)
Deliverable: a WEB-page describing the method, tests and results. Source code & binary for the predictor.
Type: Group
Credit: 10% of the grade
Design a predictor for a "two-event process" (e.g., predict whether the next card is "red" or "black"). The predictor is supposed to learn from observed data either in batch or online mode. Program the predictor and use a random-number generator to generate training and test data. Test your predictor and report
high-level description (in English) of the predictor principle,
results of the tests made with the predictor,
program source & binary for both the predictor and the generator.
No fancy interface is required. The form of the report is a WEB-page, load the template and the .gif for the bullet.
Completed projects:
Fredriksson, Himanka and Mäkeläinen
Mononen, Sadeluoto, Vasko and Mäkelä
Kontkanen, Lahtinen and Valtonen
Autio, Laczak, Lehtonen, Leivo
Project III.1 "Modeling a diagnosis problem with Bayesian networks " (deadline 5.11.97 12.00)
Deliverable: a WEB-page describing the diagnosis problem and examples of the use of the model. Hugin (.hkb) file for the corresponding Bayesian network.
Type: Group
Credit: 15% of the grade
Design a Bayesian network model of a diagnosis problem of your choice. For building the network use the demo version (5.1.) of Hugin Bayesian network software. The version for Windows 95/NT can be downloaded from here. The software is installed using the Install Wizard, you should choose the default option "Typical". The software also includes "Uninstall" - use it after you have finished your project and remove the software from the public machines. If you insist on using the Sun Sparc platform, you can download a Solaris 2.x version. Instructions and more information (introductory tutorials, examples, manual etc.) of the Hugin software can be found at Hugin home page. Test your Bayesian network with different diagnosis queries and
describe the application domain background for the diagnosis problem,
report the modeling principles you have used (what attributes were chosen, their domains, where did the probabilities come from etc.),
report a set of queries and their results
identify and discuss possible weaknesses of your model.
deliver the corresponding Hugin file.
The form of the report is a WEB-page, load the template and the .gif for the bullet.
Completed projects:
Modeling the Experiments of Baddeley's and Hitch's Working Memory Paradigm with Bayesian networks (Helin, Laine and Silander)
Modeling the Spanish Trip with Bayesian Networks (Fredriksson, Himanka and Mäkeläinen)
Improving OCR Systems with Bayesian networks (Mononen, Sadeluoto, Vasko and Mäkelä)
Modeling beer recognition with Bayesian networks (Kontkanen, Lahtinen and Valtonen)
Modeling the Parachuting Student Problem with Bayesian networks (Autio, Laczak, Lehtonen, Leivo)
Winners in F1-race (Mäkinen, Nyman and Väisänen)
Modeling the "Detection of relevant variables" problem with Bayesian networks (Hegedus)
Modeling a battle situation with Bayesian networks (Nybergh, Väliviita)
Poster on a selected topic (deadline 11.12.97 12.00)
Deliverable: poster presentation given on December 11th 1997 from 12-14.
Type: Individual or pair
Credit: 30% of the grade
Design and present a poster on a topic assigned. Copies of the poster material should be delivered to the instructor after the poster session. Some guidelines for designing the poster can be found from the following addresses: poster preparation guidelines in Chemistry, Do's and Don'ts of Poster Presentation, ASGE instructions for poster session authors, and SIAM guidelines for preparing posters.
The following is the list of poster presentation assignments
Learning Bayesian networks - an information theory based method (Autio)
Analyzing experimental psychological data with Bayesian networks and vice versa (Laine,Silander)
Probability and logic in qualitative uncertain reasoning (Laczak)
Bayesian Self-Organizing Maps (Valtonen)
Expectation-Maximization algorithm (EM) (Mäkinen)
Bayesian clustering by Autoclass (Fredriksson)
Inference ("belief propagation") in Bayesian networks by message passing schemes (Leivo)
Predictive Minimum Description Length (PMDL) principle (Hegedus)
The TETRAD approach to learning Bayesian networks (Nyman,Väisänen)
Bayes factors in model selection (Mononen)
A Bayesian approach to causal discovery (Vasko)
A Bayesian Network Analysis of a Software Process Assessment (Lehtonen)
The Dempster-Shafer theory (Himanka, Mäkeläinen)
A Bayesian Variable Selection Scheme for Classification Domains (Kontkanen)
Turbo Codes (Nyberg, Väliviita)
Applying Bayesian Networks to Mobile Agent Technology (Helin)
Project IV.1 "Empirical assessment of a model construction algorithm " (deadline 3.12.97 12.00)
Deliverable: a WEB-page describing the results achieved by the various model assessment methods.
Type: Group
Credit: 10% of the grade
Estimate the performance of a model construction algorithm with the methods discussed in class. The Naïve Bayes classifier programs (Linux platform) for building the model from a training set, and evaluating the classification results for a single test set , can be downloaded from here (Unzip & tar to create a directory URP4). The unzipped directory URP4 contains
the data file PKids.dat
README file
learner (model construction)
tester (model testing); tester returns two integers - the number of correct classifications and the total number of data items that were classified.
Your task is to try different methods for estimating the Naïve Bayes classifier performance for the "Student goals" data set (PKids.dat in the directory) - a brief description of the original (non-processed) data set is given here. The implementation of the more complex assessment methods requires some programming. Report the results for the following empirical assessment methods
internal assessment (test with training data),
train and test,
train and test subsampling,
leave-one-out cross-validation,
bootstrapping.
The form of the report is a WEB-page, load the template and the .gif for the bullet.
Completed projects:
Fredriksson, Himanka and Mäkeläinen
Mononen, Sadeluoto, Vasko and Mäkelä
Kontkanen, Lahtinen and Valtonen
Last Revised: Tuesday, 9 December 1997