From risto Thu Aug 31 18:08:08 2000 From: Risto Miikkulainen Date: Thu, 31 Aug 2000 18:07:39 -0500 To: cs381k@cs.utexas.edu Subject: class alias, notes Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin The class email alias is up and you are on it. If you send mail to "cs381k" it will go to everyone in the class, including me and Marty (the TA). We will use the alias for various announcements, questions and answers about the homework, and for general discussion about the class and AI. To see who is on the list, say "showalias cs381k". To add/chance/delete your email address from the list, send mail to martym@cs. The class notes (copies of the slides before the midterm) should be to be ready by Tuesday next week. They are available at Speedway copy, bottom floor of Dobie Mall. This year they also make them available electronically at netpaks.com. From risto Tue Sep 5 19:28:51 2000 Date: Tue, 5 Sep 2000 19:28:32 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Pointers for homework and projects Hello, everyone, First of all, if you are not in cs381k, but are receiving this message, let me know and I'll remove you from the mailing list (unless for some strange reason you'd like to follow the class). Second, if you have an account in CS, then please send me your login and email. Some of the email addresses I have now are to mail.utexas.edu, but most of you do have accounts in this department, so I'd prefer to send the email to your CS address. A few of you do not have accounts, so please do apply. As all but four of you are declared CS majors, this would be to your benefit. HOW TO DO WELL ** Start early! ** Something usually goes wrong, and late homework is penalized. ** Double-check the assignment! ** Please make sure to include everything it asks for. WRITING UP THE ASSIGNMENT Please include a plain ASCII version of all text in your report. You're welcome to include a Postscript or DVI version too if you want to use nice graphs or equations, but I'd like to have the plain text version so that I can quote the report in email when commenting on your submission. I do _not_ have MicroS*ft Word or any other word processing or spreadsheet program, so I cannot accept any documents in those formats or any format other than ASCII, PostScript, or LaTeX DVI. HOW TO SUBMIT HOMEWORK The first homework isn't due for another couple of weeks still. When you've completed it, I'd like you to write the assignment up in ASCII and submit it electronically. If you do not yet have an account on the CS machines, you still have plenty of time to get one. To submit your homework, use the /p/bin/turnin program to copy files >from your CS directory into mine. Since this program has caused some confusion in the past, I will describe it here even though a man page for it is available (type "man turnin"). The turnin program is a shell script (supported by sarvela@cs) that is used to securely copy binary or ASCII files from a directory in _your_ CS account into a directory in _my_ CS account. It runs on the CS department Linux, Solaris and AIX machines. For example, if you want to submit the tar file "hw1.tar" you would enter: /p/bin/turnin -submit martym cs381k hw1.tar This is the preferred method, as it avoids some of the problems associated with multiple submissions described below. (To make a tar (archive) file, just type (e.g.) tar cvf hw1.tar hw1 where hw1 is the directory containing the files to be put in the archive.) If you'd rather submit the three files "hw1.c hw1.out hw1.report", then you would enter: /p/bin/turnin -submit martym cs381k hw1.c hw1.out hw1.report The turnin program will list files as it submits them. If you later want to add a file "hw1.figure.ps" and update the report to fix a typo you found, you would enter: /p/bin/turnin -submit martym cs381k hw1.figure.ps hw1.report Note that the new version of hw1.report will OVERWRITE the old version, and I will not even have any record that you turned in the old version. This could then end up counting as a late submission, so it's best to make updates distinct. CONTACTING ME I am available to answer questions, either during office hours or by email. My office hours are Wed 11-12 in PAI 5.140. Office hours are appropriate for general questions requiring a lot of explanation, or just to chat about the fields of artificial intelligence, neural networks, cognitive science, and computer science. Email is more appropriate for very specific problems, particularly those related to software that I might want to check out and then reply about. Due to the large number of students in this class, please show some diligence in attempting to find the answer on your own first. If there really seems to be some problem with a system or an assignment, or if you suspect that you are overlooking or confusing something simple and important, then email me. When emailing, please include the class name (cs381k) and (if asking about a particular assignment) the assignment number (e.g. hw2) in the subject line. Finally, keep in mind that all messages to the cs381k mail alias (like this one) are archived in /u/www/users/risto/cs381k/mail-archive, which can be read using RMAIL in Emacs or, in a pinch, by typing "more /u/www/users/risto/cs381k/mail-archive". New messages do not necessarily show up there immediately. Good luck, Marty From risto Fri Sep 22 12:02:46 2000 Date: Fri, 22 Sep 2000 12:02:31 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Some Info For those of you who do not yet have accounts on department CS machines, please refer to https://www.cs.utexas.edu/users/UTCS/udb/accounts/newacct.html and follow the directions there. You will definitely need an account for the programming assignments. If you have not yet done so, please also take a look at the class web page at http://www.cs.utexas.edu/users/risto/cs381k/index.html which has the class schedule, which will have links to the homework assignments as they become ready. Also, get started early. Programming assignments tend to have a habit of taking longer than expected. Lastly, use the links at the bottom of the page. The LISP links will be very valuable for your programming. Marty From risto Thu Sep 28 16:50:21 2000 Date: Thu, 28 Sep 2000 16:50:07 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Some notes Hi, I've made some minor changes to "interface.lsp": (setq *totalmoves2* (+ *totalmoves2* 1)) as well as to some of the comments. This will not affect the functioning of your program, but is there in any case. I've also changed (setf *my-time-bound* (+ user::*time-bound* (get-internal-run-time))) to (setf *my-time-bound* (- user::*time-bound* *epsilon*)) as this is all you should need to worry about. You'll still need to determine how much time you've taken by using get-internal-run-time as well as the LISP constant internal-time-units-per-second (look at the time-associated code in "interface.lsp" to see how this is done). You can pick the epsilon that works best for you. Be sure to reference the global "user::*time-bound*" as this will be set to the same value for all programs (and may be varied for different stages of the tournament). We'll start with two minutes. Marty From risto Fri Sep 29 09:30:59 2000 Date: Fri, 29 Sep 2000 09:30:30 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: More clarifications Hi, 0. The global variables in interface.lsp are read-only. 1. You do not need to use anything from interface.lsp other than *time-bound*. However, you may use whatever is useful, if you'd like (such as the check "in-a-row"). I'll be using interface.lsp to run the tournament. So, your programs must be callable from it (i.e., you should make a .fasl file that I can load and call your "init" and "player" functions from). A good check would be to run your program against itself (or a copy of itself with a different name). 2. Your code should check your opponent's moves. You're free to use the code in interface.lsp to do so. We'll go ahead and ignore the comment in the "check" function. However, be sure that you do not change any of the global variables in the interface.lsp - only interface should change those. 3. You can use your own board representation for purposes of efficiency. You could also use "user::*board*" to see what the interface thinks is the board representation. The same applies to the *contour* variable (which is a hashtable). I hope this clarifies things. Marty From risto Fri Sep 29 13:48:41 2000 Date: Fri, 29 Sep 2000 13:48:23 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu In-reply-to: Subject: RE: More clarifications Hi, When you check for validity, all you need do is print an error message if you detect an invalid move or a false claim of victory. You can use interface.lsp as a guide. Similarly, use interface.lsp as a guide on how to track time (not the use of *init* and final). Marty From risto Sat Sep 30 14:01:59 2000 Date: Sat, 30 Sep 2000 14:01:54 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Re: Some clarifications Hi, I'd like everyone to check for validity because 1) a few programs from previous classes have made invalid moves and claims 2) your own check would help increase the probability that you do not fall prey to the same problem 3) it's not very hard and adds very little overhead Also, in my previous message, the "not" should be "note". Just want there to be no misunderstandings. Thanks, Marty > > Hi, > > When you check for validity, all you need do is print an error > message if you detect an invalid move or a false claim of victory. > You can use interface.lsp as a guide. Similarly, use interface.lsp > as a guide on how to track time (not the use of *init* and final). > > Marty > From risto Mon Oct 2 09:33:04 2000 Date: Mon, 2 Oct 2000 09:32:55 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu In-reply-to: Subject: RE: Some clarifications BTW, the homework is due Wednesday at midnight (so, Thursday and after it would start being counted as late). Marty From risto Wed Oct 4 13:01:18 2000 Date: Wed, 4 Oct 2000 13:01:13 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu In-reply-to: <39DB6F73.1CC5A4F6@cs.utexas.edu> (message from Judah Ben De Paula on Wed, 04 Oct 2000 12:57:07 -0500) Subject: Re: Displaying the board Regarding a question I've gotten about displaying the board after each move, It would be best if you made an option, say through a global variable (in your own package), such as *display-board*, which could be set to nil to turn off the display and otherwise to t to display the board. Marty From risto Thu Oct 5 15:27:44 2000 Date: Thu, 5 Oct 2000 15:27:38 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Some notes regarding HW1 submissions I have received homework from the following folks: andreasl ben bharat ftu jfan john judah liwei madhu mbilenko merugu puay rk roy rpd seth shulin sugato walkera yrchoi If you are not on this list, please review the grading policy for CS381K at http://www.cs.utexas.edu/users/risto/cs381k/homework.html. I also had described the "turnin" program in one of my first emails to the class, but there is some additional information available on the above web page. You can get to the email archives for the class from the class resource page at http://www.cs.utexas.edu/users/risto/cs381k/resources.html. Additionally, many of you did not turn in log output from your programs, although the assignment specifically asked for it. You can still submit it and I won't take off any points for it, but on future assignments, I will have to grade according to the directions. Note that this also includes the compiled "fasl" part of the assignment (made by using ":cf" in ACL per the first item in the interface specifications page). In the interest of fairness to everyone, I have to observe the late submission penalty strictly. Even if you make a small change in your program, you run the risk of overwriting your earlier (probably on-time) submission. Anything after I have to count as late. Lastly, I saw from the flurry of emails I received last night that many if not most waited until the last minute. As I'd tried to warn in an earlier email, the assignment is very likely going to take longer than expected, so it really is in your best interest to start early. Marty From risto Wed Oct 11 01:14:51 2000 From: Risto Miikkulainen Date: Wed, 11 Oct 2000 01:14:42 -0500 To: cs381k@cs.utexas.edu Subject: midterm practice questions Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Below are some sample questions that could be on the exam, just to give you an idea what kind of things I might be asking about. You'll get max benefit if you try to solve them yourself; I'll post some hints later on the toughest ones. PROBLEM SPACES What is meant by the terms ``problem space'' and ``problem instance''? Consider the task of merging several sorted lists into a single sorted list. Two lists containing $N$ and $M$ elements can be merged in $O(N+M)$ comparisons. More than two lists can be merged into a single list by repeatedly merging two lists at a time. Finding a sequence of merges that requires the lowest number of comparisons can be formulated as a search problem. For example, consider the task of merging four lists A, B, C and D. Give a state space formulation of this problem (i.e. define the problem space and instance). Illustrate by drawing part of the state space graph, with specific labels on the nodes and edges. (Note: you do not have to solve the search problem; the point is to demonstrate what the state space looks like). Give a problem reduction formulation of the problem. Again, illustrate on a partial problem reduction graph. BRUTE-FORCE SEARCH What is the asymptotic space and time complexity and solution quality of bidirectional search (justify your answer)? What restricts the applicability of bidirectional search? HEURISTIC SEARCH Consider A* with a weighted evaluation function: $f(n) = (1-w) g(n) + w h(n)$, where $0\leq w \leq 1$. Assuming that $h$ is admissible, when is the algorithm admissible (i.e.\ with what values of $w$)? GAME PLAYING In the minimax search tree below, assign values to the leaves so that $\alpha-\beta$ pruning is maximally effective for systematic left-to-right search. Indicate where pruning occurs and why. [ a picture of a search tree with 16 leaves goes here ] CONSTRAINT SATISFACTION Briefly describe two preprocessing techniques. What effect do they have on search (i.e. why do you want to do preprocessing)? SUBGOALING, MACROS, ABSTRACTION What is meant by serial and total operator decomposability? Why is it desirable? PLANNING What does the Modal Truth Criterion tell us (that is, why is it significant?) LOGIC List 2 major advantages and 3 disadvantages of using first order predicate logic as a representation formalism for AI. RESOLUTION THEOREM PROVING Resolution is much more efficient than the most straightforward implementation of Herbrand's theorem. What is the efficiency based on? From risto Wed Oct 11 23:58:13 2000 From: Risto Miikkulainen Date: Wed, 11 Oct 2000 23:58:06 -0500 To: cs381k@cs.utexas.edu Subject: hw2 Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin I've put the postscript for the first two pages of the homework asg in http://www.cs.utexas.edu/users/risto/cs381k/hw2.ps. In case you got a bad copy of the handout, you can print out those pages again. In case you didn't pick up a handout at all today, I'll put a few of them in the plastic box outside my door for now (you should have 10 pages altogether). From risto Sat Oct 14 23:37:18 2000 Date: Sat, 14 Oct 2000 23:37:14 -0500 (CDT) From: Sugato Basu To: cs381k@cs.utexas.edu Subject: Midterm discussion In-Reply-To: <200010120458.XAA22461@mamba.cs.utexas.edu> Hi everybody, How about having a discussion section before Wednesday's midterm? We could meet on Monday evening, so that if we have any doubts afterwards, we could get them cleared up by Marty or Risto during their office hours. If Monday evening sounds fine, we could meet in the Taylor basement lounge (next to the lab) at 5:00pm. -Sugato From risto Mon Oct 16 13:58:19 2000 Date: Mon, 16 Oct 2000 13:58:15 -0500 (CDT) From: Fengfeng Tu To: cs381k@cs.utexas.edu Subject: exam In-Reply-To: Is the exam open book (open notes) ? Thanks, Fengfeng From risto Mon Oct 16 14:36:28 2000 From: Risto Miikkulainen Date: Mon, 16 Oct 2000 14:36:24 -0500 To: ftu@cs.utexas.edu CC: cs381k@cs.utexas.edu In-reply-to: (message from Fengfeng Tu on Mon, 16 Oct 2000 13:58:15 -0500 (CDT)) Subject: Re: exam Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin X-Coding-System: nil Date: Mon, 16 Oct 2000 13:58:15 -0500 (CDT) From: Fengfeng Tu MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Is the exam open book (open notes) ? Thanks, Fengfeng No, it is closed book. From risto Tue Oct 17 18:34:06 2000 From: Risto Miikkulainen Date: Tue, 17 Oct 2000 18:34:02 -0500 To: cs381k@cs.utexas.edu Subject: hints on the practice questions Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin In the problem space question, the hard part is the problem reduction formulation. Here's how it goes: the top node represents the problem of forming one list out of A, B, C, and D (call it ABCD). Its descendants represent the different ways of getting that one list: (1) Merging list A with the list formed by merging B,C,D; (2) B with ACD; (3) C with ABD; (4) D with ABC; (5) AB with CD; (6) AC with BC; and (7) AD with BC. Under A-BCD you have the different ways of merging lists to get BCD (B-CD; C-BD; D-BC); and so on. In the heuristic search question, try dividing the f(n) equation with 1-w, and then observing what w would have to be for the heuristic function to be admissible. The answers to the rest of the questions come straight out of the class notes. From risto Tue Oct 17 22:05:00 2000 X-Sender: walkera@mailbox.cs.utexas.edu (Unverified) Date: Tue, 17 Oct 2000 22:03:00 -0500 To: cs381k@cs.utexas.edu From: Skip Walker Subject: Question about CSP Search In-Reply-To: <200010172334.SAA24143@mamba.cs.utexas.edu> I was wondering if anyone would like to take a stab at summing up CSP Search for me. I can't seem to remember what it's all about, and I'm making much headway understanding it from the notes. Thanks, Skip ************************************************ Skip Walker walkera@cs.utexas.edu Tetris http://www.cs.utexas.edu/users/walkera/tetris/ ************************************************ From risto Tue Oct 17 22:09:53 2000 Date: Tue, 17 Oct 2000 22:09:49 -0500 (CDT) From: Mikhail Bilenko To: Skip Walker cc: cs381k@cs.utexas.edu Subject: Re: Question about CSP Search In-Reply-To: <4.3.2.7.2.20001017220144.00abe340@mailbox.cs.utexas.edu> I found the following pages helpful in summing things up: http://liawww.epfl.ch/~torrens/Project/project/node10.html http://kti.ms.mff.cuni.cz/~bartak/constraints/constrsat.html Cheers, Misha On Tue, 17 Oct 2000, Skip Walker wrote: > I was wondering if anyone would like to take a stab at summing up CSP > Search for me. I can't seem to remember what it's all about, and I'm > making much headway understanding it from the notes. > > Thanks, > Skip > ************************************************ > Skip Walker > walkera@cs.utexas.edu > > > > Tetris > http://www.cs.utexas.edu/users/walkera/tetris/ > > > > ************************************************ > > From risto Tue Oct 24 15:56:22 2000 From: Risto Miikkulainen Date: Tue, 24 Oct 2000 15:56:19 -0500 To: cs381k@cs.utexas.edu Subject: class notes Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin The second set is ready at Speedway Copy. From martym@cs.utexas.edu Thu Oct 26 15:11:34 2000 Date: Thu, 26 Oct 2000 15:11:34 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: cs381k Hw1 Overview Reply-to: martym@cs.utexas.edu You should be getting your homework grade shortly. Avg 83.30 Min 57.00 Max 107.00 In general, a little more than half of you were above the average, but some late homeworks brought the average down, as well as a few assignments that were not turned in at all. It was pretty obvious that most of you waited until the last minute to do the homework, and so were rushed and lost points where you should have easily made them. I'll highlight some important points that will help you get more credit on the remaining assignments: 1. Include your full name and email on your reports. 2. I can't grant exceptions for late homework. Even one day late means a loss of fifteen points; two, a loss of forty. 3. The easier you make my task of grading by following the directions of the assignment, the easier it will be for me to just give you full credit. This was a very serious problem on this assignment. Only five people actually seemed to follow the directions completely. The most common problems were: (From the assignment) a) Write your program so that it can play against the user or other programs (or itself; i.e., you should be able to run (compete ``xxx'' ``xxx'') where ``xxx'' is your name). I got the impression that many of you had gotten so caught up in writing the program to be able to play against a human player that this requirement was neglected. For the next assignment, we will also need this requirement to be able to run the tournament, but we'll relax it a bit (compete ``xxx'' ``yyy''), where "xxx" and "yyy" are simply copies of your program with different package names. The specifications also asked that you give your package your last name. Most, but not all, followed this. In the next assignment, we will instead use your login as the package identifier, so please keep this in mind. b) Make the search iterative-deepening with a time bound (i.e. the program aborts the search before the time limit is exceeded and returns the current best move). The specifications say: The function "player" has to supply a move within a definite time limit which is given in a global variable "*time-bound*" (in seconds). Otherwise the player loses. Several people hard-coded the time limit in their programs. In homework 3, be sure that your time parameters are taken from the global variable above (which is accessed with "user::*time-bound*"). Also, it works to your advantage by scaling your epsilon according to the limit, rather than just making it a hard value like 5 seconds. c) Turn in some log output explaining how the program works, but turn off logging routine calls when you turn in your program for the tournament. Very few people turned in these log files initially, not to mention having sufficient explanatory logging statements. The log file should show how your program works for an entire game, with print statements of the use of various heuristics and strategies, as well as the game state. Also, most forgot to turn off the print statements (or didn't even have a toggle to do so) in the version submitted for the tournament. d) Submit a well-written report describing how the program works, Almost everyone did describe how the program worked, but a well- structured report will have sections such as "Introduction", "Program Description", "What I learned", "Major Problems", "Conclusion", etc. e) what you learned about game playing, Neglected by many, yet one of the most important aspects of the assignment (after all, you're having to go through this primarily for this reason). "I learned LISP" is not what we're looking for. f) what the major problems are for your program, and what ideas you have on how to make it play better. Again, often overlooked, but an important part of the report. For this assignment, I gave many a second chance to correct errors. So, very few lost many points on problems with the program itself. Rather, most lost points for report quality, or due to lateness. I also subtracted up to five points for not following directions. Extra points were awarded for well-written reports, particular attention to the details, and top-finishing programs in the class tournament. For homework 3, I will have to take the program as it is turned in and deduct points for each of the items listed above. With each grade report, I'm sending back the version of your program that was run in the tournament so that you have something to work from. Notice that the package definition has been commented out because I define all of the packages at once. Marty From risto Thu Oct 26 15:24:57 2000 Date: Thu, 26 Oct 2000 15:24:54 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Class Standings as of 10/26/00 ID HW1 MIDT TOTAL 7264 107.00 109.00 108.33 4470 106.00 106.00 106.00 9754 88.00 109.00 102.00 1301 101.00 99.00 99.67 0108 84.00 99.00 94.00 6368 96.00 91.00 92.67 1987 78.00 100.00 92.67 7970 81.00 98.00 92.33 0943 97.00 89.00 91.67 9974 76.00 98.00 90.67 3474 98.00 86.00 90.00 3092 78.00 95.00 89.33 1742 72.00 97.00 88.67 9197 64.00 99.00 87.33 2922 87.00 86.00 86.33 3963 79.00 83.00 81.67 1741 93.00 74.00 80.33 0339 68.00 84.00 78.67 0766 83.00 73.00 76.33 9983 85.00 71.00 75.67 5602 57.00 85.00 75.67 4409 74.00 73.00 73.33 5366 64.00 73.00 70.00 5718 0.00 72.00 48.00 1437 0.00 63.00 42.00 9860 0.00 58.00 38.67 1574 0.00 47.00 31.33 Avg 83.30 85.81 80.86 Min 57.00 47.00 31.33 Max 107.00 109.00 108.33 Std Dev 13.36 15.64 19.58 Also, you can see how the tournament went by visiting http://www.cs.utexas.edu/users/risto/cs381k/tournament.html, which also includes a link showing individual game results. -Marty From risto Fri Oct 27 16:42:18 2000 Date: Fri, 27 Oct 2000 16:42:14 -0500 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: Tournament Page Update Hi, I've updated the tournament a bit. You need to be on a CS machine to access it (or running a browser off of a CS machine). Included are directories of results from the competitions as well as of the binary compiled files if you'd like to run your program against someone else's. Marty From risto Sun Oct 29 15:34:59 2000 Date: Sun, 29 Oct 2000 15:34:55 -0600 To: cs381k@cs.utexas.edu Subject: Notes on HW2 From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hello, everyone, This is in response to some questions I've been getting: 1) Yes, you do need to turn in a report and a trace of your program on one of the examples 3-10 (cf. item 7 in the specifications handout). The report is very important. At a minimum, it should contain sections detailing your implementation, what you've learned from the assignment about resolution and implementing it, problems that your prover might have on a much more complicated proof, and what you would do to deal with these potential problems. The trace should show enough detail that one can follow the resolution process by looking at it (e.g., which clauses are being resolved and the resulting resolvent). If you are unsure as to the level of detail (as weighed against the quantity of output), just provide a "verbosity" variable that can be set to print out according to a level of detail and submit a log of what you think is reasonable. If I feel I need to see more, I can just set the variable to a higher level of detail. 2) You should be able to input the examples as they are given, and the prover should transform them into CNF, since the process is easily automated following the second page of the handout (in addition to the class notes). 3) I cannot provide solutions to the exercises as this is part of the assignment. 4) Keep in mind that the penalty for late work is strictly enforced. The assignment is due Nov 1st at midnight. Anything submitted after the time will have to be marked off. 5) Check and recheck that you've followed the directions completely. Marty From risto Sun Oct 29 17:04:16 2000 Date: Sun, 29 Oct 2000 17:04:13 -0600 To: cs381k@cs.utexas.edu Subject: Addendum From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hi, everyone, Go ahead and assume that the input clauses to your program are already in CNF. That is, you do not have to add this functionality to your program. I will award extra credit to those who do handle this, however. You can make reasonable assumptions about quantifiers and logical operators such as "forall", "some", and "-->". Marty From risto Tue Oct 31 23:31:26 2000 From: Risto Miikkulainen Date: Tue, 31 Oct 2000 23:31:23 -0600 To: cs381k@cs.utexas.edu Subject: participation points so far Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Here are the participation points so far (out of a max of 14; each time you get 0, 1, or 2 points). Looks like about half the class is participating, the other half not. Ok, let's review what's going on here: It is not the point to participate to the max each time. If everyone did that, we'd never get through the class! If you have about 5 points at this time, you are doing just fine (if you have more, you can afford to slow down and let others speak! :-). And really, if you have less than 5, please try to voice an opinion or comment or question in class more often. I'd like to hear from you. 7264 11 1301 10 9197 10 0943 9 1437 9 0108 6 4470 5 1987 5 6368 5 0766 5 7970 4 4409 3 1742 3 9983 2 3092 2 2922 2 0339 2 1741 1 3474 1 5366 0 5718 0 9974 0 3963 0 5602 0 9754 0 9860 0 1574 0 From risto Sun Nov 5 14:27:38 2000 From: Risto Miikkulainen Date: Sun, 5 Nov 2000 14:27:35 -0600 To: cs381k@cs.utexas.edu Subject: hw3 Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin I've received several questions about evolving neural nets to implement a fitness function, so let me broadcast some thoughts about it. It sounds like a good idea, and it is certainly appropriate for the assignment (that is: yes, it is fine to do it). However, in practice it has often turned out difficult to evolve networks output accurate absolute values that could be compared across different input activations. What usually works better is when you evolve the net to rank alternatives at the output. Or else, you may have to think a bit harder on what you input and output, i.e. some higher level features than just board position and its absolute evaluation value. (There is neuroevolution code e.g. in www.cs.utexas.edu/users/nn; ESP (C++) is the most powerful, JavaSANE is probably the easiest to get started with if you know java. It is possible to call C++ at least from LISP). Also, I would certainly like to see other methods used as well in this assignment. It is not very interesting if everyone does the same thing! There are many ways of obtaining supervised data, and such approaches may well be easier in the short time we have for this asg (I believe neuroevolution is a good way in the long term, but it may take weeks to really see the full benefit of such methods). For example, you could train a decision tree or neural net to "look ahead", i.e. try to predict what minimax would return, and that way extend the effective depth of the search. You could play a number of games, and collect statistics that would allow you to fix the parameters of your fitness function. All this could be done off line, using existing software (e.g. neural net or decision tree packages); only the final result needs then be imported into your lisp program, probably eliminating a lot of debugging. There are also many ways to implement learning inside your current program (which is easier in that you don't have to learn to use a simulator). Q-learning should be quite straightforward to implement in this domain (although you may want to limit the size of the board description and move choices), and I would very much like to see how it comes out. Macro learning is also a good idea, especially because it is easy to gradually improve it. So, we'd certainly like to encourage diversity of approaches. If you come up with something a little different from others, you'll get credit for it, so be daring and creative... But please start early, and start small :-) From martym@cs.utexas.edu Sun Nov 5 19:54:13 2000 Date: Sun, 5 Nov 2000 19:54:12 -0600 To: cs381k@cs.utexas.edu Subject: No office hours next week From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hello, everyone, I will be out of town next week, so I will not be holding office hours next Tuesday. However, I will still have access to my email, so hopefully I can address any questions that way until I get back next Friday. Marty From risto Mon Nov 13 19:42:15 2000 From: Risto Miikkulainen Date: Mon, 13 Nov 2000 19:42:11 -0600 To: cs381k@cs.utexas.edu Subject: class tomorrow (Tuesday) Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Just a reminder that we'll have a special session of the cs381k class tomorrow, Tuesday 11/14 7-4:45pm in BAT 215. The topic is Prolog & Logic Programming, and Rule-based and Expert Systems. From risto Wed Nov 15 19:24:36 2000 From: Risto Miikkulainen Date: Wed, 15 Nov 2000 19:24:32 -0600 To: cs381k@cs.utexas.edu Subject: the NLP homework Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin The NLP homework assignment is ready on the class web page; I also have hardcopies in the plastic box outside my door. It is due on Dec 6th. From martym@cs.utexas.edu Thu Nov 16 18:15:18 2000 Date: Thu, 16 Nov 2000 18:15:17 -0600 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: cs381k Hw2 Overview Reply-to: martym@cs.utexas.edu You should be getting your homework grade shortly. Avg 95.62 Min 58.00 Max 110.00 It looks like most of you took my last email message to heart because the reports looked much better this time, and there was much less trouble with the code. Lateness was an issue with a few, and a couple of the reports were on the skimpy side, but overall, the assignment was much better. Thanks, and keep it up. Marty From risto Thu Nov 16 18:41:52 2000 Date: Thu, 16 Nov 2000 18:41:48 -0600 To: cs381k@cs.utexas.edu Subject: Grades so far From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Here are the standings so far. "PART(ICIPATION)" is not figured in yet (which will be counted out of 14 and weighted at 5%), so the grades are a bit inflated. If you have too few points in that column, be sure to get them up before I have to add them in, too. ID HW1 MIDT PART HW2 TOTAL 7264 107.00 109.00 11.00 110.00 108.75 1301 101.00 103.00 10.00 110.00 104.25 4470 106.00 106.00 5.00 98.00 104.00 9754 88.00 109.00 0.00 110.00 104.00 7970 86.00 98.00 4.00 100.00 95.50 1987 78.00 100.00 5.00 104.00 95.50 6368 96.00 91.00 5.00 102.00 95.00 9974 76.00 98.00 0.00 106.00 94.50 0943 97.00 89.00 9.00 100.00 93.75 3474 98.00 86.00 1.00 100.00 92.50 1742 72.00 97.00 3.00 100.00 91.50 9197 64.00 99.00 10.00 90.00 88.00 0108 84.00 99.00 6.00 70.00 88.00 2922 87.00 86.00 2.00 90.00 87.25 3963 79.00 83.00 0.00 100.00 86.25 3092 78.00 95.00 2.00 76.00 86.00 0339 68.00 84.00 2.00 106.00 85.50 0766 93.00 73.00 5.00 100.00 84.75 4409 84.00 73.00 3.00 102.00 83.00 1741 93.00 74.00 1.00 91.00 83.00 5602 57.00 85.00 0.00 102.00 82.25 9983 85.00 71.00 2.00 87.00 78.50 5366 64.00 73.00 0.00 100.00 77.50 5718 0.00 72.00 0.00 76.00 55.00 1574 0.00 47.00 0.00 98.00 48.00 1437 0.00 63.00 9.00 58.00 46.00 9860 0.00 58.00 0.00 0.00 29.00 Marty From risto Tue Nov 21 11:36:59 2000 From: Risto Miikkulainen Date: Tue, 21 Nov 2000 11:35:58 -0600 To: cs381k@cs.utexas.edu Subject: reminder Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin The learning homework is due tomorrow night (11/22) at midnight. You can pick up the assignment for the last homework from my doorbox, or print it out from the class schedule page. It is due Dec 6th. No class tomorrow; have fun! From risto Sat Nov 25 17:47:30 2000 Date: Sat, 25 Nov 2000 17:47:26 -0600 To: cs381k@cs.utexas.edu Subject: Tournament From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hi, all, I've been preparing your programs for running in the tournament. Overall, they are in much better shape than before. However, I have noticed some potential problems. In several cases, programs have been submitted untrained. For reasons of fairness (as well as practical, since many of the programs depend on a lot of files for training), I cannot support training the programs while they are running in the tournament. Also, many depend on separate files (of weights or a table of Q-values) which conflict with other's files. If you submitted a program that you want to participate in the tournament as already trained, then please send a README outlining what file the program needs in order to be in the proper (trained) state, as well as the variable that turns off training. In order to compete in the tournament, I need to simply be able to have your program called through the (compete "xxx" "yyy") function in interface.lsp. Those that cannot be run through the interface will lose points for programming errors. So, in short, if your program is already trained and self- contained with *training* (or whatever you've named the toggle) turned off, you do not need to respond. Otherwise, send me a _short_ description of file(s) (one is preferable) (which I may have to rename if it conflicts with another file with the same name) that your program needs to run properly as already trained. Thanks, Marty From risto Sun Dec 3 00:18:25 2000 From: Risto Miikkulainen Date: Sun, 3 Dec 2000 00:18:21 -0600 To: cs381k@cs.utexas.edu Subject: final practice questions Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Again, here are some examples of possible exam questions, so you will have an idea what kinds of questions to expect. 60% of the questions will be about the second half, and 40% will be general questions. --------------------------------------------------------------------------- SYMBOLIC MACHINE LEARNING How would noise in the input examples affect the decision tree construction in ID3? Would you classify ID3 as a ``robust'' or ``noise-sensitive'' method compared to e.g. neural nets? NEURAL NETWORKS How does generalization depend on (a) the selection of training examples (what kind, how many...), (b) how accurately you train the network to process (noisy) training examples (i.e.\ how long you train)? REINFORCEMENT LEARNING How can value functions be learned by observing temporal differences? GENETIC ALGORITHMS AND NEUROEVOLUTION Why is crossover more efficient than mutation in producing better offspring? SEMANTIC NETS AND FRAMES How could you use a semantic network to decide the meaning of an ambiguous phrase such as ``throw up'' in a sentence? (that is, what kind of nodes and arcs would you have in the network, how would you set it up as part of the parser, and what kind of mechanism would you run on it?) NATURAL LANGUAGE PROCESSING What does ``discourse processing'' mean? List 3 issues and 3 approaches (not necessarily related). SUBSYMBOLIC AI How does distributed neural network representation of knowledge differ >from symbolic representation? Why does it make the two approaches incompatible? RULE-BASED AND EXPERT SYSTEMS Why is explanation important in expert systems? What are the two main types of queries and how are they answered? NONMONOTONIC REASONING Why is truth maintenance necessary in a nonmonotonic reasoning system? STATISTICAL REASONING How are certainty factors related to conditional probabilities? How can the rule-writer take this relation into account? ROBOTICS Briefly describe 3 approaches for a robot to determine its location. GENERAL QUESTIONS Do distributed neural network systems form a sufficient counterexample to Physical Symbol System Hypothesis (which states that symbol manipulation is necessary and sufficient for intelligence)? (here you should give arguments for and against; there is no simple correct answer). How could you use machine learning techniques to come up with a better game playing system? From risto Tue Dec 5 20:46:15 2000 From: Risto Miikkulainen Date: Tue, 5 Dec 2000 20:46:11 -0600 To: cs381k@cs.utexas.edu, cs396@cs.utexas.edu Subject: office hours Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin I cannot make it to my office hours tomorrow (Wed). If you want to see me, I'll be in my office about 6:30pm until at least 10pm (but it is best to send email and set up an appt). -- Risto From risto Tue Dec 5 20:57:02 2000 From: Risto Miikkulainen Date: Tue, 5 Dec 2000 20:56:54 -0600 To: cs381k@cs.utexas.edu Subject: review session tomorrow (Wed) Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Please bring one question on a piece of paper. From risto Wed Dec 6 16:33:01 2000 Date: Wed, 6 Dec 2000 16:32:56 -0600 From: "Marshall R. Mayberry" To: cs381k@cs.utexas.edu Subject: cs381k Hw3 Overview Reply-to: martym@cs.utexas.edu You should be getting your homework grade shortly. All in all, the programs and reports have been much better. Other than a very few oversights in some programs and only a couple of hastily written reports, I cannot really make any other suggestions for improvement. Great work, and I wish everyone the best of luck on the last homework and the final. Marty --- Hw3 Statistics: Avg 92.87 Min 46.00 Max 110.00 Std Dev 15.52 From risto Wed Dec 6 16:37:10 2000 Date: Wed, 6 Dec 2000 16:37:04 -0600 To: cs381k@cs.utexas.edu Subject: Standings From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Here are the standings as of Dec 6th. Again, "PART(ICIPATION)" is still not figured in, and the grades are accordingly a bit inflated. If you have too few points in that column, be sure to get them up before I have to add them in, too. Marty ID HW1 MIDT PART HW2 HW3 TOTAL 7264 107.00 109.00 11.00 110.00 110.00 109.00 4470 106.00 106.00 5.00 98.00 106.00 104.40 9754 88.00 109.00 0.00 110.00 106.00 104.40 1301 101.00 103.00 10.00 110.00 100.00 103.40 6368 96.00 91.00 5.00 102.00 101.00 96.20 1987 78.00 100.00 5.00 104.00 95.00 95.40 9974 76.00 98.00 0.00 106.00 96.00 94.80 3474 98.00 86.00 1.00 100.00 101.00 94.20 1742 72.00 97.00 3.00 100.00 101.00 93.40 9197 64.00 99.00 10.00 90.00 103.00 91.00 2922 87.00 86.00 2.00 90.00 105.00 90.80 0108 84.00 99.00 6.00 70.00 100.00 90.40 7970 86.00 98.00 4.00 104.00 65.00 90.20 0943 97.00 89.00 9.00 100.00 69.00 88.80 0339 68.00 84.00 2.00 106.00 102.00 88.80 3963 79.00 83.00 0.00 100.00 97.00 88.40 3092 78.00 95.00 2.00 76.00 94.00 87.60 4409 84.00 73.00 3.00 102.00 101.00 86.60 5602 57.00 85.00 0.00 102.00 85.00 82.80 5366 64.00 73.00 0.00 100.00 102.00 82.40 1741 93.00 74.00 1.00 91.00 79.00 82.20 0766 93.00 73.00 5.00 100.00 70.00 81.80 5718 0.00 72.00 0.00 76.00 95.00 63.00 9983 85.00 71.00 2.00 87.00 0.00 62.80 1437 0.00 63.00 9.00 58.00 46.00 46.00 1574 0.00 47.00 0.00 98.00 0.00 38.40 9860 0.00 58.00 0.00 0.00 0.00 23.20 Avg 84.39 85.96 5.00 95.77 92.87 83.72 Min 57.00 47.00 1.00 58.00 46.00 23.20 Max 107.00 109.00 11.00 110.00 110.00 109.00 Std Dev 13.33 15.78 3.21 12.75 15.52 20.00 From risto Mon Dec 11 12:16:55 2000 Date: Mon, 11 Dec 2000 12:16:55 -0600 To: cs381k@cs.utexas.edu Subject: Links to Tournament Results From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hi, all, I've added links to the tournament results from the CS381K homepage (http://www.cs.utexas.edu/users/risto/cs381k/index.html) for easier comparison. Marty From risto Mon Dec 11 23:34:54 2000 Date: Mon, 11 Dec 2000 21:34:49 -0800 (PST) From: puay sasiphongpairoege Subject: Group Discussion :) To: cs381k@cs.utexas.edu Content-Type: text/plain; charset=us-ascii Dear All, Madhu and I have the idea to have group discussion on 13th Dec. If anyone interested in, come to join the discussion. we shall be at taylor basement (the same place as before) around noon. Thank you all :) puay __________________________________________________ Do You Yahoo!? Yahoo! Shopping - Thousands of Stores. Millions of Products. http://shopping.yahoo.com/ From risto Tue Dec 12 01:25:38 2000 Date: Tue, 12 Dec 2000 01:25:30 -0600 (CST) From: Judah Ben De Paula To: puay sasiphongpairoege cc: cs381k@cs.utexas.edu Subject: Re: Group Discussion :) In-Reply-To: <20001212053449.10211.qmail@web1101.mail.yahoo.com> Content-Type: TEXT/PLAIN; charset=US-ASCII I'd like to but I think I'll be doing something then. Judah On Mon, 11 Dec 2000, puay sasiphongpairoege wrote: > Dear All, > Madhu and I have the idea to have group discussion on 13th Dec. If > anyone interested in, come to join the discussion. we shall be at taylor > basement (the same place as before) around noon. > Thank you all :) > > puay > > __________________________________________________ > Do You Yahoo!? > Yahoo! Shopping - Thousands of Stores. Millions of Products. > http://shopping.yahoo.com/ > From risto Tue Dec 12 12:23:15 2000 Date: Tue, 12 Dec 2000 12:23:07 -0600 (CST) From: Mikhail Bilenko To: puay sasiphongpairoege cc: cs381k@cs.utexas.edu Subject: Re: Group Discussion :) In-Reply-To: <20001212053449.10211.qmail@web1101.mail.yahoo.com> I know myself and several other people are pretty busy until Thursday because of other finals and such. I'm thinking about meeting early Friday afternoon so that we, the procrastinators, will be able to contribute something coherent to the discussion. How about Taylor basement - Friday at noon. We can also commission Puay to be an invited speaker since he'll know everything by then:) Cheers, Misha On Mon, 11 Dec 2000, puay sasiphongpairoege wrote: > Dear All, > Madhu and I have the idea to have group discussion on 13th Dec. If > anyone interested in, come to join the discussion. we shall be at taylor > basement (the same place as before) around noon. > Thank you all :) > > puay From risto Tue Dec 12 13:35:34 2000 Date: Tue, 12 Dec 2000 13:35:37 -0600 To: cs381k@cs.utexas.edu In-reply-to: Subject: Re: AI Final From: "Marshall R. Mayberry" Reply-to: martym@cs.utexas.edu Hi, all, Just verifying that the final will be this Friday, 12/15/00, in BEN 212 at 7pm. I will bring paper and a stapler. You all bring pencils or pens and ideas. Marty From risto Fri Dec 15 12:53:28 2000 Date: Fri, 15 Dec 2000 12:53:24 -0600 (CST) From: Skip Walker To: cs381k@cs.utexas.edu Subject: Re: final practice questions In-Reply-To: <200012030618.AAA18150@mamba.cs.utexas.edu> Would anyone mind sharing with me the answers to these two? Thasnks, Skpi > NEURAL NETWORKS > How does generalization depend on (a) the selection of training examples > (what kind, how many...), (b) how accurately you train the network to > process (noisy) training examples (i.e.\ how long you train)? > > SEMANTIC NETS AND FRAMES > How could you use a semantic network to decide the meaning of an > ambiguous phrase such as ``throw up'' in a sentence? (that is, what kind > of nodes and arcs would you have in the network, how would you set it up > as part of the parser, and what kind of mechanism would you run on it?) > From risto Fri Dec 15 17:53:07 2000 From: Risto Miikkulainen Date: Fri, 15 Dec 2000 17:53:04 -0600 To: walkera@cs.utexas.edu CC: cs381k@cs.utexas.edu In-reply-to: (message from Skip Walker on Fri, 15 Dec 2000 12:53:24 -0600 (CST)) Subject: Re: final practice questions Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Here's a quick suggestion towards the right direction: > NEURAL NETWORKS > How does generalization depend on (a) the selection of training examples > (what kind, how many...), You need a good sampling. (b) how accurately you train the network to > process (noisy) training examples (i.e.\ how long you train)? If you train too long your net starts memorizing instead of generalizing (i.e. it "overtrains"). > > SEMANTIC NETS AND FRAMES > How could you use a semantic network to decide the meaning of an > ambiguous phrase such as ``throw up'' in a sentence? (that is, what kind > of nodes and arcs would you have in the network, how would you set it up > as part of the parser, and what kind of mechanism would you run on it?) > This example is in the class notes. From risto Mon Dec 18 12:51:32 2000 From: Risto Miikkulainen Date: Mon, 18 Dec 2000 12:51:29 -0600 To: cs381k@cs.utexas.edu Subject: [martym@cs.utexas.edu: Re: cs381k] Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Here are the grades. If you want to pick up your final, Marty will be in his office before noon tomorrow and Wednesday afternoon. Have a good break! -- Risto ID HW1 MIDT PART HW2 HW3 HW4 FINAL TOTAL GRADE 7264 107.00 109.00 13.00 110.00 110.00 105.00 93.00 102.55 A 1301 101.00 103.00 13.00 110.00 100.00 100.00 99.00 101.35 A 4470 106.00 106.00 8.00 98.00 106.00 89.00 97.00 98.13 A 9754 88.00 109.00 2.00 110.00 106.00 104.00 97.00 97.32 A 1987 78.00 100.00 10.00 104.00 95.00 93.00 97.00 94.80 A 9197 64.00 99.00 13.00 90.00 103.00 105.00 96.00 94.60 A 6368 96.00 91.00 11.00 102.00 101.00 102.00 83.00 91.58 A 1742 72.00 97.00 3.00 100.00 101.00 97.00 95.00 90.80 A 3474 98.00 86.00 7.00 100.00 101.00 100.00 85.00 89.54 A 7970 86.00 98.00 4.00 104.00 80.00 96.00 90.00 89.24 A 0339 68.00 84.00 3.00 106.00 102.00 103.00 94.00 88.75 A 0108 84.00 99.00 12.00 70.00 100.00 93.00 84.00 88.52 A 4409 84.00 73.00 10.00 102.00 101.00 96.00 89.00 87.90 A 9974 76.00 98.00 0.00 106.00 96.00 102.00 86.00 87.70 A 3092 78.00 95.00 4.00 76.00 94.00 102.00 89.00 86.69 B 2922 87.00 86.00 8.00 90.00 105.00 91.00 81.00 85.93 B 5602 57.00 85.00 0.00 102.00 90.00 99.00 90.00 83.30 B 0943 97.00 89.00 12.00 100.00 69.00 76.00 75.00 82.87 B 3963 79.00 83.00 0.00 100.00 97.00 93.00 79.00 81.15 B 5366 64.00 73.00 1.00 100.00 102.00 100.00 84.00 80.98 B 1741 93.00 74.00 7.00 91.00 79.00 88.00 70.00 77.09 B 0766 93.00 73.00 8.00 100.00 70.00 91.00 63.00 75.13 B 5718 0.00 72.00 3.00 76.00 95.00 95.00 89.00 73.30 B 1437 0.00 63.00 13.00 58.00 46.00 79.00 85.00 65.65 C 1574 0.00 47.00 0.00 98.00 0.00 80.00 74.00 53.10 CR 9860 0.00 58.00 0.00 0.00 0.00 0.00 67.00 35.05 X Avg 84.36 86.54 7.86 96.12 93.71 95.16 85.81 83.96 Min 57.00 47.00 1.00 58.00 46.00 76.00 63.00 35.05 Max 107.00 109.00 13.00 110.00 110.00 105.00 99.00 102.55 Std Dev 13.63 15.80 4.04 12.88 14.57 7.90 9.58 14.47 From risto Mon Dec 18 12:53:37 2000 From: Risto Miikkulainen Date: Mon, 18 Dec 2000 12:53:34 -0600 To: risto@cs.utexas.edu CC: cs381k@cs.utexas.edu In-reply-to: <200012181851.MAA15854@mamba.cs.utexas.edu> (message from Risto Miikkulainen on Mon, 18 Dec 2000 12:51:29 -0600) Subject: Re: [martym@cs.utexas.edu: Re: cs381k] Reply-to: risto@cs.utexas.edu Organization: Department of Computer Sciences, UT Austin Forgot to mention: 13 was the max score on class participation (worth 5% of the total).