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

Re: PROV-ISSUE-4: agent subtypes?

From: Paul Groth <p.t.groth@vu.nl>
Date: Thu, 14 Jul 2011 20:40:19 +0200
Message-ID: <4E1F3813.3080605@vu.nl>
To: "reza.bfar@oracle.com" <reza.bfar@oracle.com>
CC: "public-prov-wg@w3.org" <public-prov-wg@w3.org>
Hi Reza, All:

I think it's important to note again that Agent is defined in a number 
of other provenance ontologies and they generally share the same 
meaning. Wikipedia says it nicely:

Agent may refer to one who acts for, or in the place of, another, by 
authority from him.

But why do we need agent for provenance. My feeling is to denote 
*responsibility*.

Who owned the Picasso then?
Who approved the publication of the document?
Who can I attribute that quote to?
Who made a particular decision?
What party made this ad in a web page?

All these use cases are about responsibility. In order for someone or 
something to take responsibility they need to be able to do things using 
their own volition. Agent is a good match for this because it usually 
refers to a person or organization, which we know have this ability to 
act on their own. When we use it in the software sense we either mean an 
intelligent agent (something or to mean a proxy for some other agent 
(i.e. user-agent in http).


In terms of trust, I would prefer that to be another discussion if 
possible. One of the insights of the Provenance Incubator Group was that 
provenance can be a platform for trust decisions but provenance should 
not be intertwined with trust.

Reza, I wonder if the notion of responsibility covers your trust use case?

Thanks,
Paul




Reza B'Far wrote:
> Satya -
>
> Thanks for clarification. Please note below (and please read reply to
> Yolanda as well).
>
>    1. I think we're talking about completely different notions of agent
>       and this is why I stressed that we shouldn't use words that are
>       outside of standard computer science or software engineering
>       vernacular. In fact, if we use those words, then we're in
>       violation of what you're saying "we become domain dependent". So,
>       I'm assuming that the only context for all the wording we use in
>       this working group is something that can be found as an
>       established understanding in the field of CS or SE at the time of
>       discussion.
>    2. My point from the papers is that Agent has an exact meaning in
>       software engineering and CS. If we mean something else, then we
>       need to find what that is in the field and use that word. If we
>       mean something completely new, then I would be VERY skeptical of
>       that NEW thing since it would indicate we're inventing stuff
>       instead of standardizing.
>    3. Again, please see the references. The word Agent has an exact
>       meaning. Fugetta, et. al, Norvig/Russell, etc. For User-Agent, you
>       can see UAProf, CC/PP, Fielding/REST, etc. below.
>
> Thanks.
>
> On 7/14/11 11:00 AM, Satya Sahoo wrote:
>> Hi Reza,
>> > Please note that the concepts of "Human agent", "System agent",
>> "Trusted agent", "Untrusted >agent", etc. have NOTHING to do with domain.
>>
>> Hypothetical scenario from biomedicine:
>> --------------------
>> A monoclonal antibody XY is introduced in an mouse model with cancer
>> to inhibit protein vascular endothelial growth factor A (VEGF-A) - XY
>> is agent involved in process of inhibition (of a protein).
>> But, it is found that XY is not effectively inhibiting VEGF-A (which
>> leads to death of the mouse).
>>
>> Upon analyzing the provenance of the XY, it is found that an error in
>> the manufacturing process of XY reduced its ability to inhibit VEGF-A.
>> --------------------
>>
>> Please identify "Human Agent" and "System Agent" in above scenario.
>>
>> Also, how does your option 2 (User Agent) cover the above scenario
>> requirements?
>>
>> The point is: each domain has its set of entities/entity states etc.
>> it identifies as "Agent". As a WG, I am not sure how do you propose to
>> enumerate all possible set of these entities/entity states?
>>
>> > I can post a half-a-dozen IEEE papers here about agents.
>> Not sure about your point here. I am sure different communities
>> (biomedical, sensor web, oceanography etc.) can post multiples of
>> dozens of papers about agent (e.g JBI, Nature, Science, JAMIA etc.)
>>
>>
>> > The usefulness of the provenance model, independent of domain, at
>> least to me, is in its ability to >accommodate the domains. Without
>> some strong typing, the standard becomes useless.
>> Agree. Modeling the provenance model in OWL will ensure strong typing
>> (i.e. it will be interpreted consistently based on OWL RDFS semantics
>> + other semantics introduced by axioms in the ontology - in your words
>> " import and export without data loss").
>>
>> Thanks.
>>
>> Best,
>> Satya
>>
>>
>> On Thu, Jul 14, 2011 at 1:13 PM, Reza B'Far <reza.bfar@oracle.com
>> <mailto:reza.bfar@oracle.com>> wrote:
>>
>>     Satya\Lena -
>>
>>     The usefulness of the provenance model, independent of domain, at
>>     least to me, is in its ability to accommodate the domains. Without
>>     some strong typing, the standard becomes useless. I state this as
>>     an opinion, but I believe this to be widely shared in the
>>     commercial software industry by practitioners. Strong typing
>>     provides a better chance of compatibility succeeding in a software
>>     market place which I believe to be of paramount importance to any
>>     standard, not just Data Provenance.
>>
>>     From an implementation perspective, and regardless of domain,
>>     unless what you're saying is to completely rule out commercial
>>     software products (again, I state, regardless of domain) which
>>     require core features such as import and export without data loss,
>>     then you have to provide some strong typing. And you can't just
>>     create some generic entity and ask people to go implement their
>>     own specific types as you will create (I would claim force)
>>     incompatibilities between different implementers which will make
>>     the standard completely useless.
>>
>>     Having said that, I'm not even sure how the domain comes into play
>>     here. Please read what I said previously, I'm going to restate it
>>     at the end of my email by re-pasting since this is something that
>>     any product implementer will feel strongly about. Please note that
>>     the concepts of "Human agent", "System agent", "Trusted agent",
>>     "Untrusted agent", etc. have NOTHING to do with domain. I can post
>>     a half-a-dozen IEEE papers here about agents.
>>
>>     In one sentence summarized - The request to consider here is to
>>     either reduce the scope of agent to something like User-Agent
>>     which is used in many other W3C standards such as HTML or to
>>     accommodate stronger types as mentioned here (and hopefully we
>>     have enough domain participants here that we can create a
>>     domain-independent sub-typed system by consensus)
>>
>>     Please read re-post below
>>     ---------------------------------
>>
>>     1. The distinction between the direct intervention of a human being
>>     effecting the state of a data versus an indirect intervention is
>>     absolutely crucial. Without this, establishing "trust" (I mean
>>     this from a formal perspective - something like PACE[1])
>>     2. I personally would lean towards one of the following options -
>>     * Strong Typing of the Agent to multiple types and specifying
>>     exactly what we mean by the types. For example, /Human
>>     Agent, System Agent/, etc. I've mentioned this in a
>>     previous thread. Within all practical usages of provenance
>>     that at least I'm concerned with, there are completely
>>     different treatments of a "snapshot" (or whatever you want
>>     to call it) of the state of an entity (which would be
>>     considered something that is included in provenance) based
>>     on whether or not there is direct human intervention (or
>>     alternatively, far more specification and strong typing) of
>>     the changes. "Agent" is way to generic to be useful
>>     practically.
>>     * Reducing the use-cases of Agent to just User-Agent which is
>>     the approach that is used in some of the other W3C standards
>>     and is weaved into the fabric of www as we know today. This
>>     would reduce the scope of what an "Agent" is. We may
>>     possibly be able to leverage work of UAProf[2] and even if
>>     not, we can learn from UAProf and CC/PP as examples.
>>     3. The key of both (1) and (2) above is that we in order to have a
>>     practical implementation, it is highly desirable to have some very
>>     exact meaning for what "Agent" is, what it does, what the boundary
>>     conditions are, etc. I also highly encourage that we do NOT
>>     include concepts that start going into RBAC and other security
>>     related standards such as Role. IMO, we need to reuse concepts
>>     from these standards.
>>
>>     I'm relatively new to the group, but have spent a lot of time
>>     reading the archives. From an implementation perspective, I
>>     caution that if things are too generic and there is not enough
>>     specification (typing) and exactness in order to accommodate a
>>     larger tent, there may be long term implementation hurdles that
>>     are presented in terms of practical implementation. In terms of a
>>     specific example, I think "Agent" above is one. It's far too
>>     generically defined at this point, IMO.
>>
>>     Please see references below.
>>
>>     [1] - PACE -
>>     http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.58.8965,
>>     http://www.mendeley.com/research/architectural-support-trust-models-decentralized-applications/
>>
>>     [2] - UAProf - http://en.wikipedia.org/wiki/UAProf
>>     [3] - CC/PP - http://www.w3.org/Mobile/CCPP/
>>
>>     On 7/14/11 9:27 AM, Satya Sahoo wrote:
>>>     Hi,
>>>     I agree with Lena here. Subtypes of agents are domain dependent
>>>     and I don't think we should define them in WG provenance model.
>>>
>>>     Regarding Reza's point, our current "definition" of agent (by
>>>     direct assertion or involved in process execution) does not (seem
>>>     to) prevent particular domain users/developers from creating more
>>>     specific sub-types of agent (s/w agent, enzymes, anti-bodies,
>>>     sensors, researcher, legal analyst etc.)
>>>
>>>     Thanks.
>>>
>>>     Best,
>>>     Satya
>>>
>>>     On Thu, Jul 14, 2011 at 12:08 PM, Deus, Helena
>>>     <helena.deus@deri.org <mailto:helena.deus@deri.org>> wrote:
>>>
>>>         Hi Luc,
>>>
>>>         I agree, agent subtypes are important. But are they not also
>>>         domain dependent?
>>>
>>>         Yes, in regard to your question  a catalyst, an enzyme are
>>>         in fact agents.
>>>
>>>
>>>         Best
>>>
>>>         Lena
>>>
>>>         *From:*public-prov-wg-request@w3.org
>>>         <mailto:public-prov-wg-request@w3.org>
>>>         [mailto:public-prov-wg-request@w3.org
>>>         <mailto:public-prov-wg-request@w3.org>] *On Behalf Of *Luc Moreau
>>>         *Sent:* 14 July 2011 16:21
>>>         *To:* public-prov-wg@w3.org <mailto:public-prov-wg@w3.org>
>>>         *Subject:* PROV-ISSUE-4: agent subtypes?
>>>
>>>         Hi Reza,
>>>
>>>         Yes, it's a good idea to discuss agent subtypes as a separate
>>>         thread.
>>>
>>>         >From my point of view, I want to be sure that we don't
>>>         disallow some kind of agents, simply
>>>         because we had not thought about them.
>>>
>>>         I believe that from a biology/chemistry point of view, a
>>>         catalyst could be seen as an agent.
>>>
>>>         Views on this?
>>>
>>>         Regards,
>>>         Luc
>>>
>>>
>>>         On 07/13/2011 07:41 PM, Reza B'Far wrote:
>>>
>>>         Graham -
>>>
>>>         Thank you for your thorough response. Please note the following:
>>>
>>>            1. I'm completely fine with sub-typing. As long as the
>>>               more concrete types (some more exact definitions of
>>>               agent) are available, I'm fine with them "inheriting"
>>>               from more generic types. My chief concern as an
>>>               implementer is to make sure that there is enough
>>>               "typing" available so that there is no loss of data in
>>>               the export/import process that can be avoided. *_So, is
>>>               the next step creation of a new email thread for
>>>               sub-typing Agent?_*
>>>
>>>
>>>         On 7/12/11 11:48 PM, Graham Klyne wrote:
>>>
>>>         Reza,
>>>
>>>         I have two main responses to your comments:
>>>
>>>         (1) your description of "Agent" here seems to me to be closer
>>>         to what the provenance work has envisaged than that described
>>>         in ws-arch document mentioned by Ryan.
>>>
>>>         (2) I fully accept your need for volitional vs computational
>>>         agent distinction for establishing certain kinds of trust in
>>>         data. But I still think that a generic agent class would keep
>>>         things simpler for developers who are not so concerned with
>>>         specific legislative or similar frameworks - I think it's
>>>         easier to subclass a generic class as needed than to unite
>>>         distinct classes.
>>>
>>>         Given that yours is a concrete use-case addressing a real and
>>>         immediate implementation need (I understand from comments by
>>>         you and your colleague) I think it may be appropriate to
>>>         include this person-vs-program distinction of agents in an
>>>         initial model, but also providing a generic agent superclass
>>>         for implementations that don't care or don't know what kind
>>>         of agent is involved.
>>>
>>>         ...
>>>
>>>         Also, I note that even in my revised understanding per your
>>>         comments, the provenance notion of "process execution" still
>>>         isn't covered by the ws-arch terminology relating to agency.
>>>
>>>         ...
>>>
>>>         You mentioned PACE. The matter of the relationship between
>>>         work in provenance and work in trusted systems came up in the
>>>         telecon to review work of the provenance incubator group, led
>>>         by Yolanda Gil. The point she made there was that [while
>>>         these are clearly interconnected] the trust work has focused
>>>         on trust in *systems*, where the provenance work is concerned
>>>         with establishing credibility in specific datasets. To this
>>>         extent, I think we need to be cautious about over-extending
>>>         the provenance model to also include concepts that would
>>>         propoerly belong in a model for trusted systems.
>>>
>>>         #g
>>>         --
>>>
>>>
>>>         Reza B'Far wrote:
>>>
>>>         Folks -
>>>
>>>         To add to Ryan's comments, I had put in a comment previously
>>>         regarding using stronger types for agents. From a practical
>>>         implementation perspective, a subset of which Ryan mentions
>>>         to be "audit" trail, etc., please note the following -
>>>
>>>         1. The distinction between the direct intervention of a human
>>>         being
>>>         effecting the state of a data versus an indirect intervention is
>>>         absolutely crucial. Without this, establishing "trust" (I mean
>>>         this from a formal perspective - something like PACE[1])
>>>         2. I personally would lean towards one of the following
>>>         options -
>>>         * Strong Typing of the Agent to multiple types and specifying
>>>         exactly what we mean by the types. For example, /Human
>>>         Agent, System Agent/, etc. I've mentioned this in a
>>>         previous thread. Within all practical usages of provenance
>>>         that at least I'm concerned with, there are completely
>>>         different treatments of a "snapshot" (or whatever you want
>>>         to call it) of the state of an entity (which would be
>>>         considered something that is included in provenance) based
>>>         on whether or not there is direct human intervention (or
>>>         alternatively, far more specification and strong typing) of
>>>         the changes. "Agent" is way to generic to be useful
>>>         practically.
>>>         * Reducing the use-cases of Agent to just User-Agent which is
>>>         the approach that is used in some of the other W3C standards
>>>         and is weaved into the fabric of www as we know today. This
>>>         would reduce the scope of what an "Agent" is. We may
>>>         possibly be able to leverage work of UAProf[2] and even if
>>>         not, we can learn from UAProf and CC/PP as examples.
>>>         3. The key of both (1) and (2) above is that we in order to
>>>         have a
>>>         practical implementation, it is highly desirable to have some
>>>         very
>>>         exact meaning for what "Agent" is, what it does, what the
>>>         boundary
>>>         conditions are, etc. I also highly encourage that we do NOT
>>>         include concepts that start going into RBAC and other security
>>>         related standards such as Role. IMO, we need to reuse concepts
>>>         from these standards.
>>>
>>>         I'm relatively new to the group, but have spent a lot of time
>>>         reading the archives. From an implementation perspective, I
>>>         caution that if things are too generic and there is not
>>>         enough specification (typing) and exactness in order to
>>>         accommodate a larger tent, there may be long term
>>>         implementation hurdles that are presented in terms of
>>>         practical implementation. In terms of a specific example, I
>>>         think "Agent" above is one. It's far too generically defined
>>>         at this point, IMO.
>>>
>>>         Please see references below.
>>>
>>>         [1] - PACE -
>>>         http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.58.8965,
>>>         http://www.mendeley.com/research/architectural-support-trust-models-decentralized-applications/
>>>
>>>         [2] - UAProf - http://en.wikipedia.org/wiki/UAProf
>>>         [3] - CC/PP - http://www.w3.org/Mobile/CCPP/
>>>
>>>         On 7/12/11 12:17 PM, Graham Klyne wrote:
>>>
>>>         Ryan,
>>>
>>>         I think the important element that is missing is that
>>>         provenance as understood so far by this group is intended to
>>>         capture actual rather than potential or unrealized processes.
>>>         This is the idea that "Process execution" aims to capture.
>>>         The notion of "Agent" as described by the ws-arch spec is, to
>>>         my mind, very much concerned with the potential rather than
>>>         the realized computation.
>>>
>>>         Although I'm not a long-time expert in this field, I think
>>>         this is quite central to the notion of provenance we're
>>>         trying to articulate and record, so it's an area where the
>>>         terminology needs to be quite distinct from other usages. You
>>>         usage of "invocation" comes closer, I think, but I'm not
>>>         convinced that yet another new term (it's not covered in
>>>         ws-arch as I recall) is helpful at this stage.
>>>
>>>         Because of the focus on actual computations, there's
>>>         correspondingly less need (or so it seems so far based on the
>>>         use-cases considered) to consider subteties of potential
>>>         processes ("Recipes", "Roles", etc.). I remain open on this,
>>>         but I would avoid adding concepts for which there is not
>>>         demonstrated need within the goals of provenance modelling
>>>         and recording.
>>>
>>>         #g
>>>         --
>>>
>>>
>>>         Ryan Golden wrote:
>>>
>>>         Thanks for taking a look at this, Graham, and I'd be
>>>         interested to hear more feedback from others. To address a
>>>         couple of your comments:
>>>
>>>         My intent with Agent was that it closely resemble the concept
>>>         of Invocation, as you say. I suppose the language "is a
>>>         computational entity" does not effectively convey the
>>>         intention. I think Invocation necessarily implies an Invoker,
>>>         so I chose a similar but broader concept of Realization. How
>>>         does does this strike you as a replacement for Process
>>>         Execution?
>>>
>>>         An Agent realizes zero or more Roles on behalf of zero or
>>>         more Persons or Organizations."
>>>
>>>         My intention with Role is to broaden the idea of Recipe to
>>>         include more abstract functions and purposes, but also to add
>>>         a subtle implication (though not requirement) that it is
>>>         something to be realized on behalf of a person or organization.
>>>
>>>         In associating Person or Organization to the concepts of
>>>         Agent and Role, the model comes closer to something that
>>>         would be useful in representing audit trails or in
>>>         establishing the trustworthiness of provenance assertions.
>>>
>>>         --Ryan
>>>
>>>         On 7/12/2011 10:00 AM, Graham Klyne wrote:
>>>
>>>         (ref. W3C Web Services Architecture Note
>>>         <http://www.w3.org/TR/ws-arch>)
>>>
>>>         Notwithstanding the slightly divergent usage in the
>>>         provenance research community, I think there is value in
>>>         using terms already adopted in the web services community
>>>         where they align - I think that would help to make our
>>>         outputs be more readily accepted, hence more relevant. Thus,
>>>         I think "Person or Organization" is reasonable term,
>>>         replacing (as I understand) what provenance efforts have
>>>         described as "Agent".
>>>
>>>         But my understanding is that "Process execution" is *not* the
>>>         same as ws-arch:"Agent", being intended to reflect a specific
>>>         invocation of the programme or service. I think the term
>>>         ws-arch:"Agent" would more closely replace "Recipe".
>>>
>>>         I'm not sure "Role" (ws-arch:"Service Role") has a direct
>>>         correspondence in the terms we've discussed to date, though
>>>         there is a notion of something like role in OPM. Similarly
>>>         for "Realizes" and "Acts on Behalf of".
>>>
>>>         #g
>>>         --
>>>
>>>         Ryan Golden wrote:
>>>
>>>         I'd like to bring a proposal up for discussion regarding
>>>         Process Execution and its related concepts. Although at the
>>>         F2F1 there wasn't much discussion over "Process Execution,"
>>>         "Generates," "Uses," and "Agent," I believe more
>>>         clarification and discussion is needed in these areas.
>>>
>>>         High Level Proposal
>>>         ----------------------------
>>>         a) Rename the concept of "Process Execution" to "Agent,"
>>>         adjusting/adding a few properties
>>>         b) Rename the concept of "Process/Recipe" to "Role,"
>>>         adjusting/adding a few properties
>>>         c) Add the concept of "Person or Organization"
>>>         d) Add the concept of "Realizes"
>>>         e) Add the concept of "Acts on Behalf of"
>>>
>>>         More Detailed Proposal
>>>         ---------------------------------
>>>         a) Concept: Agent
>>>         - is a computational entity (narrowed from "piece of work")
>>>         - may use zero or more Entity States (Bobs)
>>>         - may generate zero or more Entity States (Bobs)
>>>         - may realize zero or more Roles
>>>         - may have a duration
>>>         - may acts on behalf of a "Person or Organization"
>>>         Discussion:
>>>         Agent is a relatively well-defined industry term for an
>>>         program acting on a user's behalf. I propose it as a
>>>         replacement for "Process Execution," which has the overloaded
>>>         (and thus undesireable) term "process" in it, and does not
>>>         necessarily imply that it is acting on behalf of any one
>>>         person or organization. In scenarios involving trust, audit,
>>>         or change tracking, the ability to identify the "who" is
>>>         crucial, and so the relation between Agent and Person or
>>>         Organization is introduced. "Person or Organization" is
>>>         discussed further below. Some other common variations are
>>>         "software agent," or "user agent." One notable difference
>>>         between this concept and other agent concepts is that our
>>>         Agent may have a duration. I'm still undecided on the utility
>>>         of the duration.
>>>         There will be some discussion here about non-computational
>>>         agents. I would question the utility of being able to assert
>>>         relations involving Entity States (Bobs) and
>>>         non-computational agents, and would ask you to first consider
>>>         whether the same semantics could be better represented by a
>>>         Role instead [see next].
>>>
>>>         b) Concept: Role
>>>         - is an abstract set of tasks which pertain to a job function
>>>         - may have semantics beyond the scope of the WG model (e.g.,
>>>         as described in the RBAC reference model)
>>>         - may be realized by zero or more Agents Discussion:
>>>         Replaces the somewhat confused notions of "Agent" (as it was
>>>         discussed at F2F1), "Process," and "Recipe". Note that
>>>         multiple Roles can be realized by a single Agent.
>>>
>>>         c) Concept: Person or Organization
>>>         - is a real-world person or organization that an Agent acts
>>>         on behalf of
>>>
>>>         d) Concept: Realizes
>>>         [see Agent and Role]
>>>
>>>         e) Concept: Acts on Behalf of
>>>         [see Agent and Person or Organization]
>>>
>>>         References:
>>>         I have adapted some of this proposal from concepts in the W3C
>>>         Web Services Architecture Note
>>>         <http://www.w3.org/TR/ws-arch>, a document that I don't
>>>         entirely agree with, but which has some useful models in it.
>>>         I also referred to the NIST RBAC reference model.
>>>
>>>
>>>
Received on Thursday, 14 July 2011 18:40:57 GMT

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