From Luis.Botelho at iscte.pt Tue Oct 7 07:39:50 1997 From: Luis.Botelho at iscte.pt (Luis Botelho) Date: Tue, 7 Oct 1997 13:39:50 +0200 Subject: Emotions and AI -- Received information Message-ID: <199710071139.NAA02446@mailhost.iscte.pt> Hi, Sometime ago I've made an inquiry concerning emotions and artificial intelligence. I'm grateful to all people that has replied. Below ("Motivation and emotion: Papers, Books, Projects, Web pages, Sites and Mailing lists"), you can find all the information that has been sent to me since then. Thanks -- Luis =========================================== Luis Miguel Botelho Departamento de Ciencias e Tecnologias da Informacao Instituto Superior de Ciencias do Trabalho e da Empresa Av. das Forcas Armadas, Edificio ISCTE 1600 Lisboa 7935000, Ext. 40061 luis at iscte.pt ====================================== List of people that has replied: Josef Nerb Peter Reimann Tom Khabaza Alvaro Barreiro Garcia Frank Ritter Jeff Rickel Joyce Tang Boyland Dave Moffat Tony Kalus Paul S. Rosenbloom Phoebe Sengers Phoebe.Sengers at GS80.SP.CS.CMU.EDU Dave Cliff http://www.ai.mit.edu/people/davec/davec.html Masayuki Numao http://numao-www.cs.titech.ac.jp Scott Reilly Cristiano Castelfranchi ; JimDavies http://nis-www.lanl.gov/~jimmyd/ Wolfgang Schoppek http://www.uni-bayreuth.de/departments/psychologie Alastair Burt ===================================================== Motivation and emotion: Papers, Books, Projects, Web pages, Sites and Mailing lists Nerb, J., Spada H., & Ernst A. M. (1997). "A Cognitive Model of Agents in a Commons Dilemma" In Proceedings of the 19th Annual Conference of the Cognitive Science Society. pp. 560-565. Mahwah, NJ: Lawrence Erlbaum Associates. http://www.psychologie.uni-freiburg.de/signatures/nerb/papers/html/cogscienc e97-html.html A chapter by ROLF PFEIFFER (pp. 287-321) in: Hamilton, Vernon and Bower, Gordon H. and Frijda, Nico H. (1988) "Cognitive perspectives on emotion and motivation", Kluwer Academic Publishers Dyer, Michael G. (1987) "Emotions and their computations: Three computer models", Special Issue: Cognitive science and the understanding of emotion, Cognition \& Emotion, 3:323-347 Frijda, Nico H. and Swagerman, Jaap (1987) "Can computers feel? Theory and design of an emotional system", Cognition \& Emotion, 3:235-257 Sloman, Aaron (1987) "Motives, mechanisms, and emotions", Cognition \& Emotion, 3:217-233 Chwelos, Greg and Oatley, Keith (1994) "Appraisal, computational models, and Scherer's expert system", Cognition \& Emotion, 3:245-257 Scherer, Klaus R. (1993) "Studying the emotion-antecedent appraisal process: An expert system approach.", Cognition \& Emotion, 7(3-4):325-355 Thagard, P, & Kunda, Z. (1987).Hot cognition: mechanisms of motivated inference.Proceedings of the Annual Meeting of the Cognitive Science Society 1987, , 753-763. Damasio, A.R. (1994) "Decartes' Error: Emotion, Reason and Human Brain", Picador, London The Cognition and Affect project led by Prof Sloman Aaron Sloman http://www.cs.bham.ac.uk/~axs/ http://www.cs.bham.ac.uk/~axs/cogaff.html http://www.cs.bham.ac.uk/~axs/cog_affec/COGAFF-PROJECT.html ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/ cognition_affect at cs.bham.ac.uk The SIM_AGENT Package: Movies http://www.cs.bham.ac.uk/~axs/cog_affect/sim_demo/sim_movies.html The cognition project led by Prof. Dennet http://www.tufts.edu/as/cogstud/mainpg.htm The OZ project at Carnegie Mellon Scott Neal Reilly http://www.cs.cmu.edu/afs/cs/project/oz/web/papers.html scott at zoesis.com; (wrs+ at cs.cmu.edu - desactualizada) http://sigart.acm.org/proceedings/agents97/ (Scott Reilly Agents97 paper - PDF) http://www-cgi.cs.cmu.edu/afs/cs/project/oz/web/papers.html http://www.cs.cmu.edu/Groups/oz/papers.html http://www.cs.cmu.edu/Web/Groups/oz/ The Affective Reasoning Project Clark Elliott (DePaul University) http://condor.depaul.edu/~elliott/ http://condor.depaul.edu/~elliott/ar.html elliott at ils.nwu.edu More sites and pages http://www.ai.univie.ac.at/oefai/agents/lnai1195.html ftp://ftp.ai.univie.ac.at/papers/oefai-tr-97-02.ps.Z Emotion Research (I think this page doesn't work anymore) http://emotion.ccs.brandeis.edu/Emotion/EmoRes/CompAI/comp-refs.html http://emotion.ccs.brandeis.edu/emotion.html Affective Computing Web Page http://pendor.mit.edu/affective/affect.html Computers are Social Actors http://www.cyborganic.com/People/jonathan/Academia/Papers/Web/casa-chi-94.html Affective Computing: Home Page http://www-white.media.mit.edu/vismod/demos/affect/affect.html Agents with Faces http://tomoko.www.media.mit.edu/people/tomoko/thesis.html Organizational Measurement and Engineering: Emotion, Adaptation and Motivation http://www.ome1.com/EmotionAdaptationAndMotivation.HTML Geneva Emotion Research Group http://www.unige.ch/fapse/emotion/ The Emotional Brain http://www.cns.nyu.edu/home/ledoux/book.html Science of Emotion, The: Research and Tradition in... http://www.prenhall.com/allbooks/hss_0133001539.html Cornelius http://www.prenhall.com/allbooks/hss_0133001539.html Emotions and Emotional Intelligence http://trochim.human.cornell.edu/gallery/young/emotion.htm fungus eater http://www.unige.ch/fapse/emotion/members/wehrle/fungus.htm Dave Cliff http://www.ai.mit.edu/people/davec/davec.html http://www.cogs.susx.ac.uk/cgi-bin/htmlcogsreps?csrp434 Masayuki Numao http://numao-www.cs.titech.ac.jp http://numao-www.cs.titech.ac.jp/papers/Numao97b.ps From cl+ at andrew.cmu.edu Tue Oct 7 11:25:18 1997 From: cl+ at andrew.cmu.edu (Christian J Lebiere) Date: Tue, 7 Oct 1997 11:25:18 -0400 (EDT) Subject: ACT-R 4.0b4 Message-ID: A new beta version, ACT-R 4.0b4, is available on the ACT-R web site: http://act.psy.cmu.edu It contains a number of changes from the previous release, all of them cosmetic except for a relatively minor one. We expect this to be the last major batch of changes until the official release of ACT-R 4.0 next year. The one non-cosmetic change is that the cost parameters a and b are now learned upon failure as well as success. This seems to be the mathematically correct thing to do. Let us know if it significantly affects your model. The printing conventions have been refined. Chunk types are printed in upper case, the first letter of each word in chunk and production names is capitalized, and all other values (e.g. nil) and variables are printed in lower case. The !output! command has been generalized to make the format string optional. It can now take either a single value, or a list of mixed text and values. See the factorial example. The analogy mechanism has also been overhauled. First of all, reflecting the recent changes, it became increasingly inaccurate to call it "analogy" and we are now refering to it simply as "production compilation". Consequently, the analogy trace (:at) has become production compilation trace (:pct) and the whynot-analogy and set-analogized-parameters commands have become whynot-dependency and set-compilation-parameters respectively. Also, the dependency slot stack, which combined the functionality of the dependency slots subgoals, success and failure, has officially replaced them and those three slots have been removed from the dependency definition to try to keep down the number of slots. Finally, the variable names generated by the production compilation mechanism now reflect functionality rather than the original value in an effort to make automatically generated productions more similar to hand-written productions. These conventions were adopted for consistency with the "Atomic Components of Thought" book to be released next year. The manual will be updated soon, and in the meantime you should refer to the Notes in Patches:Release b4 (which also contains the individual patches which created this version) for details. Let me (cl at cmu.edu) know if you have any problem, question or suggestion, Christian From altmann at osf1.gmu.edu Wed Oct 1 21:27:33 1997 From: altmann at osf1.gmu.edu (ERIK M. ALTMANN) Date: Wed, 1 Oct 1997 21:27:33 -0400 Subject: ACT-R mode? Message-ID: <9710020127.AA11600@osf1.gmu.edu> Does anyone happen to have an ACT-R mode for Fred windows? Thanks, Erik. From Wolfgang.Schoppek at uni-bayreuth.de Mon Oct 6 03:58:24 1997 From: Wolfgang.Schoppek at uni-bayreuth.de (Wolfgang Schoppek) Date: Mon, 06 Oct 1997 09:58:24 +0200 Subject: ACL for Windows Message-ID: <34389A20.3156@uni-bayreuth.de> Currently I4m working with an ACL for Windows version of ACT-R 3.0 and I would like to stimulate a survey, how many of the ACT-R users use ACL. If there are some of these we should think about the best way to create an ACL version of the visual-motor components of ACT-R. Would anybody interested please drop me a note? - Wolfgang - -------------------------------------------------------------------- Dr. Wolfgang Schoppek Lehrstuhl fuer Psychologie, Universitaet Bayreuth, 95440 Bayreuth <<< Tel.: +49 921 555003 <<< http://www.uni-bayreuth.de/departments/psychologie/wolfgang.htm -------------------------------------------------------------------- From tweney at OPIE.BGSU.EDU Wed Oct 8 08:17:49 1997 From: tweney at OPIE.BGSU.EDU (Ryan D. Tweney) Date: Wed, 08 Oct 1997 07:17:49 -0500 (EST) Subject: ad (fwd) Message-ID: Position Announcement The following position announcement may be of interest to the group. The definition of "computationalist" specifically includes simulation approaches such as ACT-R! Ryan D. Tweney Department of Psychology Bowling Green State University Bowling Green, OH 43403 419/372-2301 Fax: 419/372-6013 Cognitive Psychology Position Announcement Bowling Green State University is seeking applicants for a tenure track position in computational psychology/neural networks beginning August 1998. Entry-level candidates are preferred, but all levels will be considered. Qualifications include (a) Ph.D. in psychology or related area, (b) specialization in the application of computational and/or neural network methods to problems in cognition and/or cognitive neuroscience (required) and interest in the cognitive psychology of science (preferred), (c) potential for research, publication, and external research/contract funding within the field, and (d) a strong interest in both undergraduate and graduate teaching. Responsibilities include undergraduate and graduate teaching, research and publication, external grantsmanship, supervision of master's and doctoral level students, and active participation in departmental programs and service activities. Our 32 member department offers numerous opportunities for cross-specialization collaboration and has excellent teaching and research facilities. Qualified candidates must be eligible to work in the U.S. BGSU is an AA/EEO employer and encourages applications from women, minorities, veterans, and persons with disabilities. Send vita, three current letters of reference, statements of research interests and teaching philosophy, and a transcript showing highest degree by November 28, 1997 to: Faculty Search Committee: Developmental Psychology, Department of Psychology, Bowling Green State University, Bowling Green, OH 43403. Position Announcement Cognitive Psychology - Computational/Neural Networks Department of Psychology, College of Arts & Sciences Bowling Green State University, Bowling Green, OH Closing Date for Applications All applications must be received no later than November 28, 1997. Major Responsibilities - Undergraduate and graduate teaching in cognitive psychology and general service courses to the department - Supervision of master's and doctoral level students - Ongoing program of research involving graduate students (required) and undergraduate students (preferred) - Dissemination of research in scholarly journals and professional conferences - Effort to obtain external research/contract funding for research activities and/or student training - Active participation in departmental programs and efforts and service to department, college, university, and community Qualifications - Entry-level candidates preferred, but all levels will be considered - Ph.D. in psychology or related area - specialization in the application of computational and/or neural network methods to problems in cognition and/or cognitive neuroscience (required) and interest in the cognitive psychology of science (preferred) - Demonstrated potential for research, publication, and external funding within the field - Strong interest in both undergraduate and graduate teaching - Willingness to work with faculty and graduate students to extend applications to new domains - Relevant post-doctoral experience desirable but not required Salary $38,500 for entry-level, nine-month academic appointment. Junior faculty also receive highest priority for summer teaching and research support. Effective Date of Employment August, 1998 Applications and Nominations Inquiries and nominations (which should include a letter indicating the applicant's interest in and qualifications for the position) should be directed to: Dr. Ryan Tweney, Professor, Cognitive Psychology Program, Department of Psychology, Bowling Green State University, Bowling Green, OH 43403 (ph: 419/372-8482; email: tweney at bgnet.bgsu.edu) Applications (including vita, three current letters of reference, statements of research interests and teaching philosophy, and a transcript showing highest degree) should be directed to: Faculty Search Committee: Computational Psychology, Department of Psychology, Bowling Green State University, Bowling Green, OH 43403. Bowling Green State University is an Equal Employment/Affirmative Action Employer and encourages applications from women, minorities, veterans, and persons with disabilities. Qualified candidates must be eligible to work in the U.S. Department of Psychology The Department of Psychology has five principal objectives. All are central to our primary mission. The first of these involves the offering of a broad base of undergraduate courses for educating pre-baccalaureate students from all areas of the University. Second, the Department provides strong programs for undergraduate majors, emphasizing scientific psychology as the basis and framework. It is our goal to prepare students both for employment and for post-graduate study in psychology or related fields. Third, the Department seeks to offer the highest quality graduate program possible in the broad areas of clinical, industrial-organizational, experimental (behavioral neuroscience and cognitive), developmental, and quantitative psychology and in the major subspecialties of each of these areas. Fourth, we are committed to the creation and dissemination of knowledge through research, publication, and teaching. Finally, the Department strives to meet its service commitment to the University and community. Our department includes 32 full-time faculty with approximately 100 doctoral students and over 500 undergraduate majors. Both basic and applied research interests are well represented within the department, and there are numerous opportunities for cross-specialty area collaborations among faculty and graduate students. Collectively, the faculty has published over 1000 research articles and books and has received nearly $10M in external grants and contracts. An outstanding staff and a full-time electronics technician provide friendly, professional support to the faculty. The psychology department is housed in a five-story building and provides excellent facilities for the pursuit of psychological research. Two floors (approximately 20,000 square feet) are devoted to research laboratories enhanced with state-of-the-art equipment. One research floor is devoted entirely to research with human subjects and includes laboratories devoted to the study of sensation and perception, psychophysiology, biofeedback, sleep, judgment and decision making, human-computer interactions, cognitive processes, parent-child interactions, and small group research. A second research floor houses animal research facilities with well-equipped surgical, electrophysiological, biochemical, and histological suites. General purpose rooms are available on both floors. University computer terminals in the Psychology Building provide free access to several mainframe computers, work stations, and a mail/news server. Popular software, such as BMDP, SAS, SPSS, and LISREL is available. The Institute for Psychological Research and Application, which provides outreach and research expertise to business and industry, and the Psychological Services Center, which provides mental health and training services to the surrounding area, are located in the Psychology Building. Finally, there are established relations with schools, agencies, and organizations in the Bowling Green, Toledo, and surrounding communities for field research. The quality of the Department of Psychology has been recognized by the Ohio Board of Regents Selective Excellence Program as indicated by our success in receiving a Program Excellence Award for our undergraduate program, Academic Challenge Awards for our graduate programs in clinical and industrial-organizational psychology , and an endowed chair in industrial-organizational psychology through the Eminent Scholars Program. Additional information on the department, faculty, and facilities can be found by visiting our World Wide Web site at http://www.bgsu.edu/departments/psych/. Bowling Green State University and Community BGSU is a regional, primarily-residential, state university with approximately 15,500 undergraduates and 2,500 graduate students. There are 718 full-time faculty in six academic colleges. The Department of Psychology enjoys good working relations with the College of Arts and Sciences and the university's central administration. The Bowling Green community of approximately 28,000 is a pleasant and safe environment, with a variety of housing alternatives. The City of Toledo (an SMSA of approximately 500,000) and Toledo International Airport are approximately 25 minutes from Bowling Green. Toledo offers a number of recreational opportunities including the Toledo Museum of Art, the Toledo Zoo, and the Center of Science and Industry. The cities of Columbus, Detroit, and Ann Arbor are within easy travel distances. ================================================== William K. Balzer Professor and Chair Department of Psychology Bowling Green State University Bowling Green, OH 43403 419/372-8377 419/372/6013 (fax) wbalzer at bgnet.bgsu.edu From niels at tcw2.ppsw.rug.nl Wed Oct 8 09:22:43 1997 From: niels at tcw2.ppsw.rug.nl (Niels Taatgen) Date: Wed, 08 Oct 1997 15:22:43 +0200 Subject: Parameter learning problem Message-ID: <343B8923.212711B2@tcw2.ppsw.rug.nl> Suppose I have the following set of production-rules (p retrieve-the-solution =goal> isa some-goal answer nil =fact> isa important-fact slot =answer ==> =goal> answer =answer) (p calculate-the-solution =goal> isa some-goal answer nil ==> =fact> isa important-fact slot =answer =goal> answer =answer !push! =fact) These tandems of rules are used often in many models: try to retrieve the solution, if this doesn't work, push it as a subgoal and calculate it. Now suppose we now about 50% of all the possible facts, so retrieve-the-solution will succeed 50% of the time, and calculate-the-solution takes care of the rest. Suppose further that it takes about 5 seconds to calculate the solution. So the following values apply for a,b,q and r: Retrieve-the-solution: q=0.5, a=0.2, r=1, b=0 --> PG-C = 9.8 Calculate-the-solution: q=1, a=5, r=1, b=0 --> PG-C = 15 So, in stead of trying to retrieve the fact, and if that fails calculate it, ACT-R will always calculate the fact. Since it will almost never try the other rule, it will not discover that accumulation of facts has increased the value of q. The problem is, that although the retrieve-rule might fail, you will eventually reach the goal anyway, although at a slightly higher cost. I see two possible ways solutions to this problem: - Change the parameter learning scheme for q If you do not count the failure to retrieve the fact as a failure, the extra cost of this failure will become part of b: Retrieve-the-solution: q=1, a=0.2, r=1, b=2.5 (since in 50% of the cases you still have to do a calculate-the-solution) --> PG-C = 17.3 Calculate-the-solution: q=1, a=5, r=1, b=0 --> PG-C = 15 - Change the semantics of production rules The two production rules retrieve-the-solution and calculate-the-solution very often come in pairs. The striking thing is, that the dependencies needed to learn these two rules are exactly the same, with only one difference: in the retrieve-version the fact is stored in the contraints-slot, while in the calculate-version the fact is stored in the stack-slot. So in stead of two rules, you might just want to have one rule, that first tries to retrieve the fact, and if this fails push it as a subgoal. In most of my models this would save a lot of productions! Since the second solution is rather radical, I only propose as food for thought. So I would in stead rather have a solution like the first one. Or, even better, a solution that doesn't need any change at all. So, has anyone encountered similar problems with parameters learning, and has anyone come up with a solution yet? Niels Taatgen. -- ----------------------------------------------------------------------- Niels Taatgen Technische Cognitiewetenschap / Cognitive Science and Engineering Grote Kruisstraat 2/1 9712 TS Groningen, The Netherlands Email: n.a.taatgen at bcn.rug.nl WWW: http://tcw2.ppsw.rug.nl/~niels ----------------------------------------------------------------------- From schunn+ at CMU.EDU Wed Oct 8 09:49:41 1997 From: schunn+ at CMU.EDU (Christian Schunn) Date: Wed, 8 Oct 1997 09:49:41 -0400 (EDT) Subject: Parameter learning problem In-Reply-To: <343B8923.212711B2@tcw2.ppsw.rug.nl> References: <343B8923.212711B2@tcw2.ppsw.rug.nl> Message-ID: The ubiquitous nature of the retrieve/compute pairing does make more plausible the incorporation of it into the architecture as a special case. I especially favor Niels' suggestion that the production compilation mechanism automatically created both retrieval and computation productions automatically. Let me propose a further variant. Since we have strong reason to suspect that people can use the familiarility level of the problem statement to decide whether to retrieve or calculate before trying either (see Reder & Ritter, 1992 and Schunn et al, 1997), we might think that the activation level of the goal statement could be used in conflict resolution: IF Activation(goal) > Threshold THEN retrieve, else compute The relative probabilities and costs for retrieval and computation could enter into determining the threshold. Alternatively, we could make the q for the retrieve production be a simple (logistic?) function of the activation level of the goal. This would keep the existing conflict resolution scheme basically intact. The main idea underlying both of these schemes is that the past retrieval and computations have created multiple copies of this goal chunk, and thus it will come to have a higher base-level activation. This, of course, requires that we don't modify the goal chunk when we retrieve or compute the answer, but instead create a new copy with the answer imbedded. Although this is not common practice, there have been previous suggestions that this is the correct thing to do. -Chris From ja+ at CMU.EDU Wed Oct 8 09:59:14 1997 From: ja+ at CMU.EDU (John Anderson) Date: Wed, 8 Oct 1997 09:59:14 -0400 (EDT) Subject: Parameter learning problem In-Reply-To: <343B8923.212711B2@tcw2.ppsw.rug.nl> References: <343B8923.212711B2@tcw2.ppsw.rug.nl> Message-ID: <4oCt6mO00WBO0402U0@andrew.cmu.edu> Well, thank you, Niels, for articulating the problem so well. More generally, your first solution raises the question about whether there should be a q at all. Rather, failures of the production should probably reflect in an estimate of the b parameter as you suggest. This would get rid of the one parameter and avoid the problem that our equation P = qr is in fact not usually correct. As for your second solution, it has some appeal and has been a favorite of Christian Lebiere for some time. But as you say, it is radical. Maybe a topic for discussion at the next workshop. From reder+ at CMU.EDU Wed Oct 8 10:18:00 1997 From: reder+ at CMU.EDU (Lynne M Reder) Date: Wed, 8 Oct 1997 10:18:00 -0400 (EDT) Subject: Parameter learning problem In-Reply-To: References: <343B8923.212711B2@tcw2.ppsw.rug.nl> Message-ID: Chris Schunn mentioned a couple of papers that argue that familiarity plays a role in selecting between retrieval and computation. Although I've been arguing for the role of familiarity and prior history of success as determinants of strategy choice for a very long time, John has yet to totally agree with that view. So I thought I~d give the full citations to the ones Chris mentioned and also add a few others :-) in case you are interested: Reder, L.M. (1982). Plausibility judgments vs. fact retrieval: Alternative strategies for sentence verification. Psychological Review, 89, 250-280. Reder, L.M., (1987). Strategy selection in question answering. Cognitive Psychology, 19(1), 90-138. Reder, L.M. (1987). Beyond associations: Strategic components in memory retrieval. In D. Gorfein & R. Hoffman (Eds.), Memory and learning: The Ebbinghaus Centennial Conference, Hillsdale, NJ: Lawrence Erlbaum Associates, pp. 203-220. Reder, L.M. (1988). Strategic control of retrieval strategies. In G. Bower (Ed.), The psychology of learning and motivation, Vol. 22, New York: Academic Press, pp. 227-259. Reder, L.M. & Ritter, F. (1992) What determines initial feeling of knowing? Familiarity with question terms, not with the answer. Journal of Experimental Psychology:Learning, Memory, and Cognition18, 435-451. [lead article] Miner, A. & Reder, L. M. (1994) A new look at feeling of knowing: Its metacognitive role in regulating question answering. In: Metcalfe, J. and Shimamura, A. (Eds). Metacognition: Knorwing about knowing. Cambridge, Mass: MIT Press. Kamas, E. & Reder, L. M. (1994).The role of familiarity in cognitive processing. In: E. O'Brien, and R. Lorch (Eds.),Sources of coherence in reading: A festschrift in honor of Jerome L. Myers (pp. 177-202). New Jersey: L. Erlbaum. Reder, L. M., & Schunn, C. D., (1996). Metacognition does not imply awareness: Strategy choice is governed by implicit learning and memory. In Reder, L. M., (Ed.) Implicit Memory and Metacognition. Mahwah, N.J.: L. Erlbaum, pp. 45-77. Schunn, C. D., Reder, L. M., Nhouyvanisvong, A., Richards, D. R., & Stroffolino, P.J. (1997). To calculate or not calculate: A source activation confusion (SAC) model of problem-familiarity's role in strategy selection. Journal of Experimental Psychology: Learning, Memory, & Cognition, 23, 1-27. [lead article] Nhouyvanisvong, A. & Reder, L. M. (in press) Rapid Feeling-of-Knowing: A Strategy Selection Mechanism. To appear in: Yzerbyt, V. Y., Lories, G . Dardenne, B. (1996). Metacognition: Cognitive and social dimensions. London: Sage. From lovett+ at CMU.EDU Wed Oct 8 11:40:48 1997 From: lovett+ at CMU.EDU (Marsha Lovett) Date: Wed, 8 Oct 1997 11:40:48 -0400 (EDT) Subject: Parameter learning problem References: <343B8923.212711B2@tcw2.ppsw.rug.nl> Message-ID: I will pipe in too since Chris (Schunn) just alerted me to these emails. I have definitely encountered this conflict-resolution situation and can offer one easy practical solution: changing the value of G (from the default 20) changes the trade-off between costs and accuracy. For example, with G=10, the difference in q's is de-emphasized and the two productions have PG-C of 4.8 and 5. In this situation, the noise would be enough to get the system to try retrieving (long enough to see that q eventually > 0.5) Niels's first solution (changing the parameter learning scheme for q) is interesting. It highlights the possibility that one can always design a model that emphasizes cost-learning over probability learning. Also, note that the current interpretation of q for retrieval productions is relatively new, so there may be some backlash on the P=q*r equation. The second solution (and Chris S's first suggestion) are radical in that they imply retrieval always preceds computation. A slightly less radical version might be to include in models the combo rule Niels describes (if retrieval fails push a compute goal) along with a competing compute rule. This way, one could short-circuit to computing w/o having first failed at the retrieval. Finally, Chris's notion of incorporating the "familiarity with the problem" effect on these conflict-resolution issues is one that has been around. His idea of allowing goals that are just new versions of previously held goals have more activation is interesting. It means changing the semantics of goal activation though. One could imagine that the activation of a goal is a JOINT function of its familiarity (as in chunk practice) and the attention directed toward it (as in some amount of W). This would produce effects like the following: it is easier to "attend" to a familiar goal (requires less W to make the goal useful for other retrievals). --Marsha From tj at earth.medinfo.ohio-state.edu Wed Oct 8 12:29:11 1997 From: tj at earth.medinfo.ohio-state.edu (Todd R. Johnson) Date: Wed, 08 Oct 1997 12:29:11 -0400 Subject: Parameter learning problem In-Reply-To: References: <343B8923.212711B2@tcw2.ppsw.rug.nl> Message-ID: <3.0.1.32.19971008122911.00a779f0@earth.medinfo.ohio-state.edu> I've had similar problems with the compute/retrieve rules. One potential source of the problem is that the retrieve rule bypasses recognition. One possible solution might be to attempt to recognize the goal prior to deciding between compute/retrieve. This would be easy to do with two rules--one that recognizes the goal and a second fallback rule that indicates no recognition. The retrieve/compute rules would be conditioned on the results of recognition. This technique might also make the Act-R models more consistent with the feeling of knowing data. While testing experimental software to run alphabet arithmetic problems, I noticed that I often recognize previously seen problems, even though I cannot always recall their solution. In all cases, recognition seems to preceed recall. I also found that I would sometimes begin to count up the alphabet, while "simultaneously" attempting to retrieve a solution to a recognized problem. By "attempting to retrieve" I mean that I have a "feeling that I am trying to retrieve" while counting up the alphabet. Sometimes I retrieve the answer and am able to stop counting. Given the importance of compute/retrieve in many Act-R models, it looks like its time to seriously model the feeling of knowing results. From cl+ at andrew.cmu.edu Thu Oct 9 15:26:47 1997 From: cl+ at andrew.cmu.edu (Christian J Lebiere) Date: Thu, 9 Oct 1997 15:26:47 -0400 (EDT) Subject: 1998 ACT-R Summer School and Workshop Message-ID: After yesterday's provocative discussion (finally, the mailing list working both ways!!), I have a more pedestrian topic. The leaves are starting to turn colors here in Pa, which means that our thoughts are starting to turn to next year's summer school and workshop. The summer shool format is by now pretty well oiled, with two five-day weeks separated by a free weekend. Even though it has occasionally been suggested to decouple the summer school from the workshop, summer school students greatly benefit from the research exposure. The length of the workshop has evolved somewhat, but it is generally agreed that a workshop shorter than 3 days would either be too hurried or too restrictive, and a workshop longer than 4 days would be too demanding. And of course the last half day should be subtracted to allow for travel, leaving us with either 2 1/2 or 3 1/2 days. Evidence from this year suggests that we can generate enough material to fill 3.5 days, but increased selectivity may allow to cut it to 2.5 if desired. As practiced this year, mornings would be devoted to about 6 half-hour research presentations, and afternoons would feature special-interest panel discussions and tutorial sessions. As suggested, a relatively informal schedule of demos would be included on the evening of the second and perhaps the third day. Currently, the main question concerns the date. The summer school must start after June 15 to accomodate European students. August is generally considered a poor month, with many people on vacation. Moreover, we have always been intent on avoiding conflict with the 4th of July weekend and the CogSci conference, scheduled this year for (correct me if I am wrong) July 31 to August 4. This essentially leaves us with 4 possibilities: Summer School Workshop 1) June 15 to 26 June 27 to 30 2) June 29 to July 10 July 11 to 14 3) July 6 to 17 July 18 to 21 4) July 13 to 24 July 25 to 28 1) is the earliest possible date. 2) schedules the free summer school weekend on the 4th of July. 4), like this year, leaves two free days between the workshop and cogsci. This schedule was designed to ease the travel burden on our foreign contingent, but many people complained about being away from their families for too long and not having time to prepare for cogsci. 3) would be a compromise, leaving 9 days between the workshop and cogsci. Let us know how you feel about the schedule and other issues. If you would simply like to register your opinion, you should send email directly to me (cl at cmu.edu). If you would like to raise an issue that warrants discussion, feel free to reply to the group (act-r-users at andrew.cmu.edu). Thank you for your feedback, Christian From conzalez at andrew.cmu.edu Mon Oct 13 12:01:55 1997 From: conzalez at andrew.cmu.edu (Cleotilde Gonzalez) Date: Mon, 13 Oct 1997 12:01:55 -0400 (EDT) Subject: No subject Message-ID: Christian, I would like to go to the ACT-R workshop during Summer 98. How can I find more information about it? Also, Javier Lerch told me there is a lisp tutorial available. I would like to know how to get it. thank you, Cleotilde Gonzalez CMU-GSIA Center for Interactive Simulations/ Human-Computer Interaction Institute 5000 Forbes Ave. GSIA room 226 Pittsburgh PA 15213 tel: (412) 268-6242 fax: (412) 268-5063 http://www.andrew.cmu.edu/user/conzalez/ From niels at tcw2.ppsw.rug.nl Thu Oct 16 09:07:12 1997 From: niels at tcw2.ppsw.rug.nl (Niels Taatgen) Date: Thu, 16 Oct 1997 15:07:12 +0200 Subject: Parameter learning problem Message-ID: <34461180.CF6DEB95@tcw2.ppsw.rug.nl> I want to thank everyone for their contributions to the discussion. As I now look at it, there are two problems to be solved: - The current parameter learning mechanism does not take into account the fact that after a failed production the goal may still be reached. - How can the "feeling of knowing" be modeled in ACT-R and applied to the compute/retrieve choice. With respect to the first problem, I wondered why we make destinctions between q and r, and a and b? Why not estimate P and C directly from experience, and use these as production parameters? Since q and r are always multiplied, and a and b are always added, I cannot think of a good reason to estimate them seperately. Maybe there are good reasons for this, but it's my guess that there is still some part of ACT-R 2.0 in here, where q,r,a and b themselves were calculated from effort, r*, b* and whatever. So a possible, attractive solution with respect to parsimony, is to use P instead of q and r. (And, although it currently poses no direct problems, use C instead of a and b). For the second problem I have no clear intuitions. Todds solution sounds appealing, although I am not sure what he means by "Recognizing the goal". It sounds like it can be done with the current architecture, which is of course always more attractive than requiring changes. Niels. -- ----------------------------------------------------------------------- Niels Taatgen Technische Cognitiewetenschap / Cognitive Science and Engineering Grote Kruisstraat 2/1 9712 TS Groningen, The Netherlands Email: n.a.taatgen at bcn.rug.nl WWW: http://tcw2.ppsw.rug.nl/~niels ----------------------------------------------------------------------- From reder+ at CMU.EDU Thu Oct 16 09:34:16 1997 From: reder+ at CMU.EDU (Lynne M Reder) Date: Thu, 16 Oct 1997 09:34:16 -0400 (EDT) Subject: Parameter learning problem In-Reply-To: <34461180.CF6DEB95@tcw2.ppsw.rug.nl> References: <34461180.CF6DEB95@tcw2.ppsw.rug.nl> Message-ID: <0oFVTMq00iWU02_nk0@andrew.cmu.edu> Excerpts from mail: 16-Oct-97 Parameter learning problem by Niels Taatgen at tcw2.ppsw. > For the second problem I have no clear intuitions. Todds solution sounds > appealing, although I am not sure what he means by "Recognizing the > goal". It sounds like it can be done with the current architecture, > which is of course always more attractive than requiring changes. Perhaps what Todd means is that ACT-R should be modified to allow more data driven behavior rather than being so "top down." John is very goal oriented and as the theory stands it may be a good model of him, but he's lived with me long enough to know that there exist people that are much more stimulus driven. Couldn't there be a high level production that notices when elements in the environment are very active? When something is very active perhaps one tries to match it to a goal. As for the case of feeling of knowing: When there is a goal to answer a question, perhaps the production that tries to answer the question could begin by evaluating the activation of the problem representation to decide whether to search (or continue search) as opposed to always trying search before calculating the answer. I think there is a lot of evidence to support the psychological validity of that process (I'll spare you the personal citations this time). --Lynne From ja+ at CMU.EDU Thu Oct 16 09:30:57 1997 From: ja+ at CMU.EDU (John Anderson) Date: Thu, 16 Oct 1997 09:30:57 -0400 (EDT) Subject: Parameter learning problem In-Reply-To: <34461180.CF6DEB95@tcw2.ppsw.rug.nl> References: <34461180.CF6DEB95@tcw2.ppsw.rug.nl> Message-ID: <4oFVQFK00iVD01ya00@andrew.cmu.edu> Good questions again The reason for separating q from r and a from b is to enable the rG - b discounting of subgoals. The idea here is that a subgoal is only worth as much as the goal is after the goal has been achieved. Thus, we need the downstream quantities. I think in 2.0 we used to assign PG-C as the value to the subgoal but this was too severe since some of the P and some of the C was associated with the subgoal. I still like the idea, however, of eliminating the q parameter and letting just a represent the cost of the production (and potential subgoal it spawns). Then the discounting of the subgoal becomes effectively PG - b since q = 1. For example, suppose q = 1 (proposal) r = .6 a = 10 b = 2 G = 20 The old proposal would have assigned a value of .6 * 20 - 10 - 2 = 0 to the subgoal and it would be immediately abandoned. The current proposal would assign a value of .6*20 - 2 = 10 since that is the expected value of the goal after the subgoal has been completed.