<div> Mention of CYC reminds me that Ken Forbus has been doing a lot of work with CYC to automatically "read" and encode inputs to the kinds of analogical reasoning processes that he and Dedre Gentner do, in order to remove the element of subjectivity that has been in the encoding process.</div>
<div> </div>
<div>Susan Chipman<br><br> </div>
<div><span class="gmail_quote">On 6/24/08, <b class="gmail_sendername">Ball, Jerry T Civ USAF AFMC AFRL/RHAT</b> <<a href="mailto:Jerry.Ball@mesa.afmc.af.mil">Jerry.Ball@mesa.afmc.af.mil</a>> wrote:</span>
<blockquote class="gmail_quote" style="PADDING-LEFT: 1ex; MARGIN: 0px 0px 0px 0.8ex; BORDER-LEFT: #ccc 1px solid">Troy,<br><br>A few years ago we looked at integrating CYC with ACT-R (caveat: we never actually implemented anything). It turns out that the basic form of knowledge representation in CYC (logical predicates and assertions) is quite different from what is supported in ACT-R (chunks --aka frames-- and productions). In fact, early CYC used frames, but was later converted to logical predicates. Automatically mapping CYC logical predicates and assertions to ACT-R chunks and productions is a non-trivial task, but certainly one worth pursuing.<br>
<br>More recently, with respect to language modeling, I have come to believe that it is important to have a good idea of the range of tasks an ontology will be used to help perform. Creating ontologies in the abstract, without reference to some task or tasks, is very open-ended and doesn't directly facilitate modeling task behavior. This may explain, in part, why CYC has not had more success. The ontology used in the language model is motivated by the need to process various grammatical constructions. New categories in the ontology are introduced, as needed, to fulfill functional requirements. In the long run, this approach may (or may not) lead to an ontology which is generalizable to new tasks, but in the short run it is important to have the ontological concepts needed to perform the task at hand. Of course, the more complex the task, the more complex the ontology, and language comprehension is certainly a complex task, ultimately requiring a fully general ontology, but one that is functionally motivated.<br>
<br>Jerry<br><br>-----Original Message-----<br>From: <a href="mailto:act-r-users-bounces@act-r.psy.cmu.edu">act-r-users-bounces@act-r.psy.cmu.edu</a> [mailto:<a href="mailto:act-r-users-bounces@act-r.psy.cmu.edu">act-r-users-bounces@act-r.psy.cmu.edu</a>] On Behalf Of Emond, Bruno<br>
Sent: Tuesday, June 24, 2008 10:50 AM<br>To: Kelley, Troy (Civ,ARL/HRED); Susan Chipman; Dan Bothell<br>Cc: <a href="mailto:act-r-users@act-r.psy.cmu.edu">act-r-users@act-r.psy.cmu.edu</a><br>Subject: Re: [ACT-R-users] ACT-R robot (UNCLASSIFIED)<br>
<br>Troy,<br>This looks like we a sliding into another thread.<br>Anyhow, you might want to look at the WN/Lexical as another model for<br>implementing large conceptual models. There is a parameter in WN/lexical to<br>load Wordnet chunks either in Declarative memory or in the WN/Lexical<br>
module. You might be able to reuse some of the WN/Lexical code.<br>Bruno<br><br><br><br>On 6/24/08 1:25 PM, "Kelley, Troy (Civ,ARL/HRED)" <<a href="mailto:tkelley@arl.army.mil">tkelley@arl.army.mil</a>><br>
wrote:<br><br>> Classification: UNCLASSIFIED<br>> Caveats: NONE<br>><br>> Susan,<br>><br>> We have been looking at using a semantic network as a starting point<br>> for declarative memory chunks. There are two major semantic networks that<br>
> we are looking at using. ConceptNet from MIT and OpenCYC which is a spin<br>> off of Doug Lenat's work at the University of Texas. We are still working<br>> on exactly how to interface a semantic network with declarative memory<br>
> chunks, but I think the general idea has promise. We are also trying to use<br>> a production syntax so that the same general productions don't have to be<br>> developed over and over. Both of these ideas have promise but we are still<br>
> working on the implementations.<br>><br>> Troy D. Kelley<br>> AMSRD-HR-SE<br>> Army Research Laboratory<br>> Human Research and Engineering Directorate (HRED)<br>> Aberdeen Proving Ground, Aberdeen MD 21005-5425<br>
> voice: 410-278-5869<br>> fax: 410-278-9523<br>><br>> -----Original Message-----<br>> From: <a href="mailto:act-r-users-bounces@act-r.psy.cmu.edu">act-r-users-bounces@act-r.psy.cmu.edu</a><br>> [mailto:<a href="mailto:act-r-users-bounces@act-r.psy.cmu.edu">act-r-users-bounces@act-r.psy.cmu.edu</a>] On Behalf Of Susan Chipman<br>
> Sent: Tuesday, June 24, 2008 1:10 PM<br>> To: Dan Bothell<br>> Cc: <a href="mailto:act-r-users@act-r.psy.cmu.edu">act-r-users@act-r.psy.cmu.edu</a><br>> Subject: Re: [ACT-R-users] ACT-R robot<br>><br>> As the former sponsor of much of the ACT-R work, I would really like<br>
> to have seen a cumulative library of model building work in order to avoid<br>> duplication of effort and the associated waste of scarce research funding.<br>> In particular, it seemed to me that declarative knowledge should not be<br>
> built over and over again. I retired, so I am no longer in a position to<br>> push this agenda, but I hope the community will come to such a conclusion<br>> soon.<br>><br>> Susan Chipman<br>><br>><br>
> On Tue, Jun 24, 2008 at 12:00 PM, Dan Bothell <<a href="mailto:db30@andrew.cmu.edu">db30@andrew.cmu.edu</a>> wrote:<br>><br>><br>><br>><br>> --On Tuesday, June 24, 2008 8:04 AM -0700 Bruce J Weimer MD<br>
><br>> <<a href="mailto:bjweimer@charter.net">bjweimer@charter.net</a>> wrote:<br>><br>><br>>> Also, I was wondering as I read the papers, is information that's<br>> learned<br>>> during a session with the robot stored or saved when the robot's<br>
> turned<br>>> off or is it lost? I'm reading through the ACT-R tutorials and<br>> it seems<br>>> that each time a model is run, it's reset to run again - there<br>> (so far)<br>>> hasn't been any mention of a mechanism to save the information in<br>
>> declarative memory, weights, latencies, etc.<br>>><br>><br>><br>> There is no mechanism built into ACT-R for saving the internal state<br>> of any<br>> module or the system in general. However, depending on what you<br>
> want to<br>> save, it's usually not too difficult to just cache that out and read<br>> it<br>> back in later.<br>><br>> For example, here's a function which will create a file with the<br>> commands<br>
> to rebuild the chunks in declarative memory and restore their<br>> parameters<br>> (only one of the declarative parameters is included here for<br>> brevity):<br>><br>> (defun save-chunks (file-name)<br>
> (let ((chunks (no-output (dm)))<br>> (params (no-output (sdp :name :reference-count)))<br>> (cmdt (car (no-output (sgp :cmdt)))))<br>><br>> (sgp-fct (list :cmdt file-name))<br>><br>> (command-output "(add-dm ")<br>
><br>> (dolist (x chunks)<br>> (command-output "(")<br>> (pprint-chunks-fct (list x))<br>> (command-output ")"))<br>> (command-output ")")<br>><br>> (dolist (x params)<br>
> (command-output "(sdp ~A :reference-count ~s)" (first x)<br>> (second x)))<br>><br>> (sgp-fct (list :cmdt cmdt))))<br>><br>> The file that writes out could then just be loaded later to restore<br>
> the<br>> model's declarative memory. A similar function could be written to<br>> dump<br>> the productions and their parameters (see the reference manual for<br>> the<br>> relevant commands).<br>><br>
> Depending on how the general parameters for the model are set there<br>> may<br>> be some other things that would need to be adjusted as well, but<br>> those<br>> sorts of things are going to vary from model to model. One example<br>
> would<br>> be either advancing the time to where the "old" model left off or<br>> adjusting<br>> all of the declarative parameters to reflect the "current" time<br>> being<br>> necessary if base-level learning were enabled.<br>
><br>> Dan<br>><br>> _______________________________________________<br>> ACT-R-users mailing list<br>> <a href="mailto:ACT-R-users@act-r.psy.cmu.edu">ACT-R-users@act-r.psy.cmu.edu</a><br>> <a href="http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users">http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users</a><br>
><br>><br>><br>> Classification: UNCLASSIFIED<br>> Caveats: NONE<br>><br>> _______________________________________________<br>> ACT-R-users mailing list<br>> <a href="mailto:ACT-R-users@act-r.psy.cmu.edu">ACT-R-users@act-r.psy.cmu.edu</a><br>
> <a href="http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users">http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users</a><br><br>--<br>Bruno Emond. Ph.D.<br>Research Officer | Agent de Recherche<br>Tel. | Tél. 1.613.991.5471<br>
Facsimile | Télécopieur 1.613.952.0215<br><a href="mailto:bruno.emond@nrc-cnrc.gc.ca">bruno.emond@nrc-cnrc.gc.ca</a><br><br>Institute for Information Technology<br>Institut de technologie de l'information<br><a href="http://iit-iti.nrc-cnrc.gc.ca">http://iit-iti.nrc-cnrc.gc.ca</a><br>
<br>National Research Council Canada<br>Conseil National de Recherches Canada<br>1200 Montreal Rd., M50, Ottawa, ON K1A 0R6<br>Government of Canada | Gouvernement du Canada<br><br><br><br><br>_______________________________________________<br>
ACT-R-users mailing list<br><a href="mailto:ACT-R-users@act-r.psy.cmu.edu">ACT-R-users@act-r.psy.cmu.edu</a><br><a href="http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users">http://act-r.psy.cmu.edu/mailman/listinfo/act-r-users</a><br>
</blockquote></div><br>