W3C home > Mailing lists > Public > public-prov-wg@w3.org > July 2011

Re: Agent Sub-Types

From: Graham Klyne <GK@ninebynine.org>
Date: Thu, 14 Jul 2011 22:53:00 +0100
Message-ID: <4E1F653C.7010005@ninebynine.org>
To: reza.bfar@oracle.com
CC: public-prov-wg@w3.org
Reza,

I don't know about minority, but I think something you said is very important to 
keep in mind: "However, if the group is saying that defining touch points to the 
tangent layers, per your own references, is also out of scope, then I warn that 
there is a fundamental problem for product implementers."

IMO, it's crucial that we articulate those "tough points", even if we don't 
define any notion of agent subclassing _within_ the provenance model.

#g
--


Reza B'Far wrote:
>    Ok.  At this point, I'm resigning to the fact that I'm in the 
> minority.  However, just look at HTML as a data point:
> 
> They could have made <p></p> and <div></div> tags into the same generic 
> tag with an attribute (or attributes).  I would argue that they made the 
> right decision(s) in the strong types they created, though not all were 
> perfect decisions (e.g. <blink>).  That "strongly encouraged" all the 
> implementers of browsers to define rendering of a paragraph in a 
> specific way (and different from <div) which is more loose).  The core 
> issue is compatibility.  If you make something generic and provide only 
> soft guidelines for extensions, you end up with compatibility issues 
> when product implementers build their products because they will do 
> whatever they can within the specification to differentiate their 
> products.  Ignoring that fact risks success of a standard once implemented.
> 
> But, I end with my original statement that I'll drop the topic.
> 
> On 7/14/11 1:57 PM, Daniel Garijo wrote:
>> Hi Reza, all,
>> I agree with you in that subtyping is very important from an 
>> implementer persepctive.
>> However I think that the model discussed in the model TF is supposed 
>> to be generic,
>> and once we have it, the test cases TF can develop some profiles 
>> subtyping all the
>> generic concepts, showing examples of how it would be extended in 
>> different domains.
>> Thus, developers could use these profiles as reference for other 
>> extensions.
>> Best,
>> Daniel
>>
>> 2011/7/14 Reza B'Far <reza.bfar@oracle.com <mailto:reza.bfar@oracle.com>>
>>
>>     I agree that it's the right thing to do to keep trust definition
>>     out of the scope of WG.  However, if the group is saying that
>>     defining touch points to the tangent layers, per your own
>>     references, is also out of scope, then I warn that there is a
>>     fundamental problem for product implementers.  If you want to
>>     exclude all aspects of trust including any think like the ability
>>     to embed something else via a URI or something like that to a
>>     trust mechanism, then you'll have compatibility issues from
>>     different product vendors.  If you want to do that knowingly, it's
>>     fine and I'll drop the thread, but if you disagree and think that
>>     exclusion of trust doesn't cause fundamental incompatibility, we
>>     can continue thread and I can provide more details on why this is
>>     the case.
>>
>>     So, my point from the beginning is that without subtyping, things
>>     are too generic to be able to import and export things about
>>     entities between different systems.  And I believe a primary
>>     use-case for usage of the model is import/export between different
>>     implementations.
>>
>>
>>
>>     On 7/14/11 12:35 PM, Satya Sahoo wrote:
>>>     Hi,
>>>     I agree with Yolanda that core of provenance should not include
>>>     trust, since in view trust is a function of provenance (computed
>>>     over provenance assertions). In a paper by Sizov et al. [1],
>>>     provenance is modeled as a layer between trust and proof layers
>>>     of the Semantic Web layer cake. 
>>>
>>>     Some comments on Reza's point:
>>>     > for the first version, we need something that the implementers
>>>     can provide that says "the person >creating this mod is not
>>>     trusted" or "the person creating this mod is trusted" at that
>>>     binary simplicity >level.
>>>     A follow up query would be (in context of provenance) - "why is
>>>     the person trusted or not trusted". Is it due to the algorithm
>>>     used to compute trust (there are several, e.g. [2] [3]) or is it
>>>     the provenance of the person or the provenance of the mod (which
>>>     provides the context for trust)?
>>>     In addition, how is the trust value in the above statement
>>>     represented - binary value, a plain text label, a term from a
>>>     trust vocabulary/ontology? 
>>>
>>>     Hence, I believe trust is not in scope of the WG.
>>>
>>>     Best,
>>>     Satya
>>>
>>>     [1]
>>>     http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4397215&tag=1
>>>     <http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=4397215&tag=1> 
>>>     [2]  H.Luo andJ.Tao and Y.Sun Entropy-BasedTrustManagementforData
>>>     Collection in Wireless Sensor Networks, Proceedings of WiCom 09.
>>>     5th International Conference on Wireless Communications,
>>>     Networking and Mobile Computing,  page(s): 1-4, 2009.
>>>     [3] Y. Wang and M.P. Singh. Formal Trust Model for Multiagent
>>>     Systems. In Proceedings of the 20th International Joint
>>>     Conference on Artificial Intelligence (IJCAI-07). pp. 1551 -
>>>     1556, 2007.
>>>
>>>     On Thu, Jul 14, 2011 at 2:00 PM, Reza B'Far <reza.bfar@oracle.com
>>>     <mailto:reza.bfar@oracle.com>> wrote:
>>>
>>>         Yolanda -
>>>
>>>         Thank you for the response.  Please see responses below -
>>>
>>>            1. You're completely correct that trust has shades of gray
>>>               (accuracy, preciseness, etc.).  This is partly why I
>>>               also included the PACE reference.  However, it should
>>>               be up to the implementer to determine trust.  All we're
>>>               doing is providing some very coarse grain way to even
>>>               express existence or lack of trust.  Perhaps we should
>>>               add to the two that I put in an "Unknown".  At this
>>>               point, IMO, for the first version, we need something
>>>               that the implementers can provide that says "the person
>>>               creating this mod is not trusted" or "the person
>>>               creating this mod is trusted" at that binary simplicity
>>>               level.  Later on, during future versions of the draft,
>>>               additional attributes can always be added.  I'm even
>>>               find with doing that now... or creating a pointer to
>>>               other standards that deal with trust.  But, not dealing
>>>               with it makes it so that the fact that an agent is
>>>               mentioned is not all that useful if I have to have
>>>               trust.  And most, if not all, commercial applications
>>>               have to have trust.  It's not an option.  I can't go
>>>               republish some news from some random source that I
>>>               don't have any trust for or no one vouches for as a
>>>               reputable org (journalism use-case).  Nor can I provide
>>>               records management lineage in time for some legal
>>>               evidence piece. 
>>>            2. I am fine with the proposal of completely removing
>>>               agent.  I guess it's better than ONLY having a
>>>               "generic" agent.  But I prefer specific agent(s) 
>>>            3. References from Fugetta, et. al, as well as
>>>               Russell\Norvig, Taylor/Dashofy, Medvidovich etc. where
>>>               Software Agents are definitively defined look at the
>>>               following categories -
>>>                   * Mobile Agents - mobility context
>>>                   * Intelligent Agents - automated processes that
>>>                     make their own decisions without direct human
>>>                     interaction
>>>                   * User-Agent as defined in Http/HTML/etc. within
>>>                     the context of client-server computing
>>>            4. On (3) above, my "beef" here is that we need to use
>>>               words that have definitive meaning in software
>>>               engineering within their own context.  System Agent is
>>>               typically used (and I previously sent a reference on
>>>               this) to refer to automated intelligent agent... some
>>>               cron job that's running in the background doing
>>>               automated stuff.  User-Agent is defined by Fielding in
>>>               REST. 
>>>            5. Orthogonal to discussion - I generally don't like
>>>               something called "recipe" for example.  I mean what is
>>>               a recipe?  It's in my kitchen, but I don't find it in a
>>>               gang-of-four software engineering book or in anything
>>>               that I've seen in a graduate or undergraduate software
>>>               engineering book.  Getting creative with words is
>>>               dangerous.  And I don't think we're inventing anything
>>>               here in this (or any other) working group in the way of
>>>               a new theory, principle, etc. so I strongly recommend
>>>               we use exact words that are in either accepted and
>>>               semi-mature (few publications, not just 1 paper) or
>>>               fully mature computer science and/or software
>>>               engineering disciplines.
>>>
>>>         Best.
>>>
>>>
>>>         On 7/14/11 10:40 AM, Yolanda Gil wrote:
>>>>         Hi Reza:
>>>>
>>>>         You raise an interesting topic, albeit a tough one.
>>>>
>>>>         Trust tends not to be binary, it comes in all shades of grey
>>>>         (e.g., a degree of confidence).
>>>>
>>>>         It is also subjective, the level of trust may depend on the
>>>>         application, the domain, or the use of the provenance.  
>>>>
>>>>         So in my opinion, the core of a provenance representation
>>>>         should not include a representation of trust.  Maybe later
>>>>         we include an extension to represent trust, but note that
>>>>         many trust metrics can be derived from a given provenance
>>>>         record.
>>>>
>>>>         I am also not sure about your second category.  I am not
>>>>         sure if the NYT as publisher of an article would be
>>>>         considered "user-agent" or "system".  I am not sure if my
>>>>         personal email agent should be considered "system" or
>>>>         "user-agent".
>>>>
>>>>         In general, I think ontologizing agency is tricky.  
>>>>
>>>>         In my opinion, the notion of agent should be eliminated from
>>>>         the model unless we want to attach a special meaning to a
>>>>         participant which is a meaning of responsibility for a
>>>>         step/process.
>>>>
>>>>         Yolanda
>>>>
>>>>
>>>>
>>>>         On Jul 14, 2011, at 10:18 AM, Reza B'Far wrote:
>>>>
>>>>>         Creating new thread to put agent sub-typing up for discussion.
>>>>>
>>>>>         Proposal is to have the following sub-types of agent
>>>>>
>>>>>            1. Trust-based sub-types
>>>>>                   * Trusted Agent
>>>>>                   * Untrusted Agent
>>>>>            2. Limiting the scope of System vs. Human interaction
>>>>>                   * User-Agent
>>>>>
>>>>>         Alternative to 2, we could also do Automated System Agent
>>>>>         and Human Agent.
>>>>>
>>>>>         Reza
>>>>
>>>
>>
Received on Thursday, 14 July 2011 21:55:16 GMT

This archive was generated by hypermail 2.2.0+W3C-0.50 : Thursday, 26 April 2012 13:06:37 GMT