August 20 (Thu), 1998 Introduction [J]: Rule-Based Systems: Chapter 1 [J]
August 25 (Tue), 1998 Syllabus and Errata Sheet handed out Introduction [J]: Rule-Based Systems and Their Deficiencies: The Myth of Modularity, Truth-Functionality, Detachment, Bidirectional Inference, Explaining Away, Abduction and Deduction.
August 27 (Thu), 1998 HW #1 assigned: exercises 2.1, 2.2, 2.3, 2.4, due Tuesday. Qualitative Issues: Causal Networks. D-separation. Intro to Probability Theory for Bayesian Networks. [J, Ch.1]
September 1 (Tue), 1998 HW 1 discussed; will be picked up on Thursday. A little more probability. Classical (Laplace), frequentist, and subjective (de Finetti) interpretation. The Dutch Book Theorem: sketch with an example. Bayes' theorem (or rule: the inversion formula). Use of Bayes's rule in diagnosis: illustration from Charniak and McDermott's 1985 text. References shared: Geiger, Dan, Thomas Verma, and Judea Pearl. "Identifying Independence in Bayesian Networks." _Networks_, 20 (1990), 507-534. Neapolitan, Richard. _Probabilistic Reasoning in Expert Systems_. Wiley, 1990. (The _Networks_ paper may be a good one for graduate student presentation.)
September 3 (Thu), 1998 HW 1 collected. Correction given in class, with extensive discussion. Independence and conditional independence. Definition of Bayesian network (from Neapolitan).
September 8 (Tue), 1998 HW 2 assigned: exercise 2.5 due Thursday. Chain rule for Bayesian networks (Theorem 2.1 [J]). Two proofs given. The second one is based on topological sorting. (Two or three students raise their hands when I ask who has not heard of topological sorting.) Icy roads example revisited, with numbers.
September 10 (Thu), 1998 HW 2 collected: Exercise 2.5. HW 3 assigned: Exercises 2.6 and 2.7, due Tuesday. Cluster trees: illustration using icy roads and sprinkler examples; direct computation of joint probability for the burglary example. Ch. 3 started: dog out example, including some discussion of quantitative issues. Graduate student Neville Yoon is assigned to present BOBLO.
September 15 (Tue), 1998 HW 3 was not collected; some background material was discussed. It will be collected on Thursday. HW 4 assigned: Exercises 2.8 and 2.9, due Tuesday, September 22. Neville Yoon presents BOBLO, to everyone's satisfaction! Introduction to Hugin, in the lab (SUM 330). (In the afternoon, e-mail to Finn Jensen sent, with request for results of BOBLO validation.)
September 17 (Thu), 1998 HW 3 collected. HW 4 confirmed. HW 5, due Tuesday, September 22, is exercise 3.1 (which requires use of Hugin). Various examples of "capturing the structure." The stud farm example, with parameters.
September 22 (Tue), 1998 HW 4 collected. Extension given for HW 5, which will now also include Exercise 3.2. Homework assigned: HW 6: 3.3, 3.4, 3.8, 3.13 due Tuesday, September 29 HW 7: 3.5, 3.6, 3.15 due Thursday, October 1 HW 8: 3.16 for graduate students only, due Thursday, October 1. Further discussion of simple Bayes models. More on "capturing the structure." The transmission of symbol strings example (3.2.3).
September 24 (Thu), 1998 HW 5 (exercises 3.1 and 3.2) collected. HW 6 modified: 3.13 is not included in it; 3.13 will be due later. Batch learning, using transmission of symbol strings as a running example (3.4.1, 3.4.2, 3.4.3, 3.4.4).
September 29 (Tue), 1998 Midterm.
October 1 (Thu), 1998 Correction of Midterm. Faithfulness of DAGs to independence models: discussion using part (i) of telescopes exercise in Russell and Norvig. Verma and Pearl's theorem on equivalence of Bayesian networks: skeletons and V-structures; mimicking; Occam's razor (minimality).
October 6 (Tue), 1998 Completion of Ch. 3: modeling tricks (Relations, Noisy OR, Independence of Causal Influence, Divorcing), adaptation, discussion of exercise 3.3 and (briefly), 3.8, 3.13. HW 6 modified again: 3.3, 3.8, 3.13 due Thursday, Oct 15.
October 8 (Thu), 1998 HW 6 collected (3.3, 3.8, 3.13) HW 7 modified and assigned, due next Thursday, Oct 15: 3.7, 3.15; for graduate students only: 3.16. For 3.15, we assume that spoon and knife are used for continental breakfast; fork and knife for British breakfast. Readings assigned, to be presented by graduate students on Tuesday, Oct 20: MUNIN (IJCAI 87 and 89 by Yoon and Shi); CHILD (AIS-93 and _Models from Data_ (1994)). Ch.4 started (propagation in Bayesian networks): intro, 4.1 (algebra of belief tables), 4.2 (probability updating in joint probability tables), 4.3 (cluster trees) up to and not including Absorption in Cluster Trees.
October 13 (Tue), 1998 Fall Break
October 15 (Thu), 1998 HW 7 collected. Continuation of Ch.4.
October 20 (Tue), 1998 Correction of exercises 3.7 (daughters) and 3.8 (Monty Hall). Presentation of MUNIN by graduate students (Yoon and Shi).
October 22 (Thu), 1998 HW 6 (Exercises 3.3, 3.8, 3.13) returned. HW 8 assigned: exercises 4.1, 4.2, 4.3, 4.4, due Thursday, October 29. Presentation of CHILD (Blue Babies) by graduate students (Johnson and Turkett) Continuation of Ch.4: proof of Theorem 4.6 (running intersection property implies that the marginal on V of T(U) computed at any cluster that includes V is the same.
October 27 (Tue), 1998 Completion of Section 4.5 up to and including the statement of Theorem 4.10.
October 29 (Thu), 1998 HW 8 (4.1 through 4.4) collected. Theorem 4.9 (graph is eliminatable iff triangulated) explained and proved in full (!). Demos of breakfast Hugin programs will be given next Tuesday, unless canceled by me.
November 3 (Tue), 1998 Election day--no classes. Bill Turkett presents his breakfast Hugin program.
November 5 (Thu), 1998 HW 9 assigned: 4.6, 4.8(iv only), 4.9. Theorem 4.10 (connected graph is triangulated iff it has a junction tree) explained and proved in full (!). Explanation (using visit to Asia as example) of Theorem 4.11 (spanning tree of a junction graph is a junction tree iff it is a maximal spanning tree) and of the construction of junction trees from elimination ordering described in exercise 4.8. (Note that proof of Theorem 4.11 was given at October 28 DAG Journal Club meeting.)
November 10 (Tue), 1998 HW 9 collected; no new HW assigned. Stochastic simulation and the Gibbs sampler. Exercises 4.13 (on reachability and reducibility), 3.16 (on the NP-hardness of probability computation), and 4.14 (on the NP-hardness of finding a configuration of positive probability, done in class. Reference for 3.16 is: Cooper, Gregory F. "The Computational Complexity of Probabilistic Inference Using Bayesian Belief Networks." _Artificial Intelligence_, 42 (1990), 393-405.
November 12 (Thu), 1998 Computing marginals from first principles. Interaction graphs and Non-Serial Dynamic Programming.
November 17 (Tue), 1998 The method of cutset conditioning: exercise 4.11 (parts 1-5). Value of Information: Section 5.5 and drHugin paper (half).
November 19 (Thu), 1998 HW 10 assigned (due Tuesday, Nov. 23): exercises 4.11 (parts 1-6) and 4.13. Value of Information: Section 5.5 and drHugin paper (almost completed).
November 24 (Tue), 1998 HW 10 collected (exercises 4.11 1-6 and 4.13). HW 11 assigned: Exercise 5.13 (Hugin exercise). Papers assigner to graduate students: Heckerman and Horvitz (UAI-98) for Turkett and Yoon, to be presented on Tuesday, December 1; Heckerman, Horvitz, et al. (UAI-98) for Johnson and Shi, to be presented on Thursday, December 3. Conclusion of DrHugin paper on VOI. Computation of joints: Section 5.1, but especially computation by modifying the clique tree (a.k.a. junction tree). Data conflict: presentation of Kim and Valtorta's paper from UAI-95.
November 26 (Thu), 1998 Thanksgiving. No classes.
December 1 (Tue), 1998 Presentation of Heckerman and Horvitz paper by Turkett and Yoon. Influence diagrams in Hugin: apple grower example. Demonstrations of Hugin exercise 5.13 scheduled.
December 3 (Thu), 1998
Presentation of Horvitz et al. paper by Shi and Johnson.
Assignment of take-home exam:
CSCI 509--002 Fall 1998
Final Exam
Assigned Thursday, 98/12/3
This is a take-home exam.
Your task is to do exercise 6.2 of the textbook, using Hugin.
You must carry out this exercise on your own without consulting any source
other than the printed textbook and your class notes. In particular, you are
not allowed to consult the demontration disk and any web resources.
You will demonstrate your influence diagram between 9 and noon on Thursday,
December 10 (the regular final exam time), according to a timetable to be
determined at our last class meeting. If you have been asked to make
modifications to the programs you wrote for other assignments, you will have to
demostrate those programs too. As you demonstrate your programs, I
may ask you questions related to all material covered in the course.
In order to pass the course, you must have turned in all homework by Tuesday,
December 8, at 5pm. This is a ***firm*** deadline. All assigned homework is
recorded in the log file accessible through the course web page.