From furm@npac.syr.edu Fri Nov 19 10:32:58 1999 Date: Thu, 18 Nov 1999 13:50:28 -0500 (EST) From: Wojtek Furmanski To: Tom Pulikal , Krishnan Rangarajan , Hojung Lim , Zeynep Odcikin Ozdemir Cc: Wojtek Furmanski Subject: DARPA Agents Here are some entry points into the DARPA Agents space. Please check and see if you can find some more info relevant for the proposal. Proposal is DARPA ISO 00-07 on Agent Based Computing - see http://www.darpa.mil/iso/ABC/BAA0007PIP.htm. Teknowledge wants to bid for DAML Integration task. This PIP contains a link to CoABS (http://coabs.globalinfotek.com) which is an onging DARPA agents program. Other useful links in this PIP include SHOE (http://www.cs.umd.edu/projects/plus/SHOE/), OML (http://wave.eecs.wsu.edu/CKRMI/OML.html) and On2broker (http://www.aifb.uni-karlsruhe.de/WBS/www-broker/). SHOE is very important since it was led and also largely developed by Jim Hendler while on Univ of Maryland faculty (http://www.cs.umd.edu/users/hendler/) and he now leads DAPRA Agents (you may find some of his talks in CoABS site). One interesting person in current CoABS team is Craig Thompson from Objects Services and Consulting, Inc. - see his project page http://coabs.globalinfotek.com/coabs_public/project_overview_pages/ agentility_objs.html where he talks about Java+CORBA+XML+HLA based agents testbed and his info page at http://www.objs.com/agility/index.html. He also leads now the OMG Agent Working group (http://www.objs.com/isig/agents.html) - see in particular their Green Paper (ftp://ftp.omg.org/pub/docs/ec/99-10-02.doc) and OMG RFI at http://www.omg.org/techprocess/meetings/schedule/Agent_Technology_RFI.html ftp://ftp.omg.org/pub/docs/ec/99-07-01.htm. According to Teknowledge, Hendler wants to make DAML the next big thing on the Internet i.e. some sort of a cosmic standard that is used by everyone like current browsers and search engines. As for current agents standards, they are being collected and managed by FIPA (http://www.fipa.org/). In particular, for Meta-Ontology Service, they are adopting OKBC (http://www.ai.sri.com/~okbc/spec.html) which was also developed by DARPA. However, Teknowledge does not like OKBC as you can see in the attached message and they would rather use CYC (allegedly the largest knowledge base, started by Teknowledge and contined by a spinoff) and the associated language MELD - see links below. I also include some strategy comments by Rick Hayes-Roth, Teknowledge CEO, that might be useful as a guide in searching for relevant materials. The focus should be on military and on simple, established ontologies. So I'm thinking of using DMSO ontologies for M&S such as CMMS, OML etc. and applying it for computer generated forces i.e. agents for training etc. --------------------------------- >From apease@teknowledge.com Thu Nov 18 13:21:30 1999 Date: Wed, 17 Nov 1999 13:32:41 -0800 From: Adam Pease To: Wojtek Furmanski Cc: Wojtek Furmanski , lcoker@teknowledge.com, John Connell , Rick Hayes-Roth Subject: Re: (A real answer) Re: NSF ITR, DARPA Agents Wojtek, In my view (danger, soapbox talk incoming...), the problem with most agent work to date is that they've focused on interchange formats or negotiation protocols or message syntax etc. They haven't focused on content. If we're to create real agents that have interesting interaction I believe they need extensive knowledge and semantics. They'll have to share a large body of semantics in order to communicate effectively, just as humans need to share an expressive grammar (natural languages) along with a huge amount of shared context (semantics) about the physical world, cultural norms, people's behavior etc. So, I think we should focus on the content that others have ignored. In order to allow for rich content we should choose an expressive language like CycL/MELD or KIF. OKBC packs features of an implied language (frames) into the protocol, thereby limiting the usefulness of the protocol. There's a KIF spec at http://logic.stanford.edu/kif/ , MELD is described at http://www.cyc.com/cycl.html Adam At 12:55 PM 11/17/99 -0500, Wojtek Furmanski wrote: > >Hi, Adam, >I just noticed OKBC as recommended in the FIPA Meta-Ontology Service >and so I decided to explore it it more detail. I can't comment on >scalability as I didn't try it yet (but I just noticed some public Java >implementation from Stanford/SRI and I was planning to check it out). >So far I just went through the spec and I noticed some overlap >with MOF and UML meta-model by OMG so I thought these two approaches >could be cross-examined and integrated somehow. Regarding limitations >of frame-based reps, I remember your words for IDIAM and comments >on your MELD language - would you view MELD as a better entry point >for DAML? Where could I read something more about it? >Speaking of FIPA, they also got something called ACL (Agent >Communication Language) which seems to be natural for XML-ization >and it admits multiple ontologies and 'content languages' (including >W3C RDF) so I thought this could be a useful entry point as well. >I am now also looking into OML, SHOE, CG, On2Broker etc. and getting more >and more confused.. >Wojtek > >On Wed, 17 Nov 1999, Adam Pease wrote: > >> Wojtek, >> I'd like to get some detail from you on what you think about OKBC. Under >> the HPKB program it was only really used by Stanford and SRI. We found >> that it doesn't scale well to large KBs, and it's tied to a limited frame >> representation. It might be used for inspiration in defining some new >> protocol for DAML but I'd initially advise against using it directly. What >> do you think? >> >> Adam > > > > ____________________________________ Adam Pease apease@teknowledge.com Teknowledge (650) 424-0500 x571 -------------------------------------------- >From rhayes-r@teknowledge.com Thu Nov 18 13:23:06 1999 Date: Wed, 17 Nov 1999 14:07:22 -0800 From: Rick Hayes-Roth To: Adam Pease , Wojtek Furmanski Cc: Wojtek Furmanski , lcoker@teknowledge.com, John Connell Subject: Re: (A real answer) Re: NSF ITR, DARPA Agents I gave Jim Hendler a couple of talks to read when he was putting together his BAA. One of them was a talk on "Why Computers Can't Help Me." It was a simple talk for an interdisciplinary audience, but the essential message is: If you want an agent to help you and be able to direct it in few words, it has to understand your context very well. The more context-free its understanding, the less likely it is that it will be powerful enough to do much of interest. Even a dog needs to know what you mean by "paper" when you tell it to fetch the paper (i.e., it's the wrapped up thing on the driveway that smells like newsprint). In any case, I think the trick is to look for the sweet spot, where moderate amounts of knowledge which is fairly stable can be used to mediate contextualized requests that enable requestors to get big bang for the bits (return on effort to communicate). In essence: the goal is high value-added assistance, especially for the kinds of workers intended to justify the program (in this case intelligence analysts and military decisionmakers, who do a lot of situation analysis and briefings (as do lots of other people)). Rick At 01:32 PM 11/17/99 -0800, Adam Pease wrote: >Wojtek, > In my view (danger, soapbox talk incoming...), the problem with most >agent work to date is that they've focused on interchange formats or >negotiation protocols or message syntax etc. They haven't focused on >content. If we're to create real agents that have interesting interaction >I believe they need extensive knowledge and semantics. They'll have to >share a large body of semantics in order to communicate effectively, just >as humans need to share an expressive grammar (natural languages) along >with a huge amount of shared context (semantics) about the physical world, >cultural norms, people's behavior etc. > So, I think we should focus on the content that others have ignored. In >order to allow for rich content we should choose an expressive language >like CycL/MELD or KIF. OKBC packs features of an implied language (frames) >into the protocol, thereby limiting the usefulness of the protocol. > There's a KIF spec at http://logic.stanford.edu/kif/ , MELD is described >at http://www.cyc.com/cycl.html > > >Adam