OMG Agents Working Group

Minutes of Meeting #8

Mesa, Arizona

Co-chairs:  Stephen McConnell, James Odell, Craig Thompson

OMG Agent WG homepage:  http://www.objs.com/isig/agents.html

internet/00-01-02

Agenda

Attendance

Minutes

by Craig Thompson

Review of Relevant Other OMG Work

Asynchronous Messaging, Bill Binko, TRC, Bill.Binko@trc.com

Orbos-98-05-05 or current RTF activity. Several have read the spec. The spec extends CORBA with AMI - asynchronous messaging. Client calls server and server response comes back to client message handler. Spec adds priorities, store and forward, timeouts, queueing or router, possibly several of these. Client and server can be disconnected. In AMI, it redoes the stubs - IN parameters are passed in GIOP. Standard form of message. This is type safe, not manually setting up the mapping. DII is not type safe. You can support guaranteed delivery, priorities, but not everything you can do with MOM. Now client sends f() to stub, goes to client orb to server orb to POA to skeleton to call f() on server side. Main change is on client side to talk to reply handler that is skeleton for servant. Callback model. Client side ORB must be messaging aware. Server side does not have to change (in theory). When you can get failure (and it is not all in memory). Francis comments that these are like continuations. You could make this transparent but if you want control over the synchrony, then you need to know the handler exists and be able to control it. For instance, an indirection is to have a router in an intermediary process then this lets client or server to go away. You can specify policies based on the intermediaries. Apparently, the claim is that transactions do not work over the intermediaries. Priorities across many clients cannot be balanced. Many intermediate routers are not ORBs at all but might be MQSeries. These intermediaries are not true ORBs in that they do not support all ORB interfaces. The unshared transaction model is a hack. Client can call transactional services but not be part of a transaction. For queuing, the router builds a value type that is passed back to client that can check status of the request via querying the value type. Generally IIOP talks to router but router can talk anything to other routers that must eventually talk to Corba client. Most ORBs will support lower level proprietary extension for passing messages not in IIOP. All compliant ORBs are required to put the route into the IOR. ORBs can use more dynamic routing but they must support this static chain idea.  Right now, messaging and interoperable security do not work well together.  All this should be in JIDL, Iona, Visigenics 3.0 ORB.

So how is all this relevant to agents and ACL? Type safety might be something agents do not need, thinks Francis. Agent message is more like inter program email. The email is highly structured. You could package your xml acl encoded message as an ANY and reply with a Null. Our routers could be smart. Could also steal notification work as well (though that is broadcast). Agents could be built on the AMI and notification design patterns.

Action item: Portable Interceptor Spec - get updated on this. Can add new service profiles this way.

On-line Upgrades, Bill Wood, CMU/SEI, wgw@sei.cmu.edu

Real-time PSIG is working on RFP for online software updates. Want continuously operating systems even when there are updates. Some proprietary software does this. Not doing load-balancing - (another RFP is doing this. Want to solve the interface versioning problem but backed off due to ORBOS advice. Types of upgrades, different implementation, QoS or semantics. Non-redundant objects - either you queue requests -or- deny service. Redundant objects - emphasis is here. Current fault tolerance work ongoing. There are active and passive kinds of fault tolerance. Passive has downtime versus more complexity but better response. Replications within the group is not enough for configurations. An element of preserving system state. Rollback to an old version. Selection mechanisms like first arrival, voting, preferred, decision making and safety, update paradox (functional redundancy so have same signature but different implementations, and all the same where if one fails they all will). Another approach is to install second version simultaneous and check both logs to see if they behave the same. Went through mandatory requirements and optional that included how to safely add a new member into a group.

Problems with AB - fault tolerance is going on now so we need to get some experience. So is single entity upgrades only worth doing? Agent WG definitely thinks it is.  What kinds of versioning? (branching or linear). Also must affect interface upgrading? (not sure). Q: what about dependencies?  Not covered in this RFP.

CSI Interoperable Security RFP, Ed summary in lieu of real presentation

We were hoping for a briefing on this RFP but will try again at the next meeting.  A few people gave us some inputs.  The SEC SIG is moving toward using an SSL type of approach. Vendors did not like SEC IOP so this avoids that.  So you could not get to security level 2.  Permits out of band communication. Has nothing to do with portable interceptors. Possibility exists to have single security service that is portable across ORBs.

Agent groups needs to define what kinds of non-repudiation, integrity services, signing, etc. Can you seal data and carry it around and know it has been tampered with. What about multi-hop authentication. What about delegation? What happens when agent security policy is different than host security policies. What are agent security requirements?  One answer is, they are not too unique and different agent systems might need different capabilities.   Tune in again at the next meeting.

Registration and Discovery RFP, Steve McConnell

Steve goes over the RFP.  This RFP was presented to AB, Telco task force, and EC task force. Many on the internet are producing XML schemas. XML Schema does have inheritance. It is version 2 of XML, more of less.  These are schemas describing things that have commercial terms and conditions. The proposed Registration and Discovery Service is tied very closely to Trader. Trader itself can store non-IDL, just URL or IOR. XMI is an XML based schema for exchanging UML across repositories. Includes a mapping of all IDL types. Has all the semantic info.

Selection languages include XQL and commerce net selection language among others. Conjunctive and disjunctive, multiple discovery services, predicates, natural language keywords, financial products markup language query spec,

Trying to avoid management and plugin subsystems, hard to define this.

How do you avoid the ontology problem of resource descriptions describing anything? Answer is to define to a collection of policies for versioning, terms and conditions, .

Submitters LOI due 10 March.

Architecture Board pointed out there are no requirements in the management area. Want to include rationale for why not to include it as a non-goal. Add some scalability criteria. Steve says:  not trying to provide replaceable systems, just interoperable ones.

Discussion on the meaning of policy, which is way underspecified in the RFP.

More below.

FIPA News, Fancis McCabe

No FIPA meeting since last OMG meeting; next one in ten days. Lots of changes in FIPA management. Everything is changing, specs management, meeting format, architecture of specs, spec of architecture. Recent work on FIPA architecture. Presentation by Kate Stout. Reminiscent of three RFP track - naming, registration, transport. Three things the architecture group is working on now.

A main thing in the future is policy. Want predicate calculus like language to describe policy. To use this service you must register with that service first. You must use secure message transport. Management of policy. Policy domain is a set of agents agreeing to a certain set of policies. Highly controversial. Question about point of application of policy. A statement about a constraint on application of agent. Applied at point of use of service. Another cut on notion of domain is as a computational vehicle and agent pre-agreeing to policies. Policy verification is a membership test. Similar to client server relationship in an agent-service relationship. Policies are related to services.

Francis wrote a paper on Abstract ACL.

Lots of discussion in FIPA on content languages. Requirements on them, policies on them. Do you allow multiple representations of content languages.

Can an entity have a set of object messages (so it is an object) and also send and receive ACL messages (so it is simultaneously an agent).  Provides a good way for objects and agents to interoperate in the same environment.  Another way to do this is via a content language that is oo.  Some agents may not also be able to speak IDL. This leads to a picture in which agents and objects are both first class and the same object may (or may not) be both an agent and an object, depending on whether it is speaking ACL or IDL.

We need a compelling set of use cases for ACLs. FIPA uses SL as a content language but it is undecidable. Content language in FIPA is plugin affair. Ask prolog query or Tell prolog assertion. But what happens when a prolog rule contains another ACL query?

Agent Technology Green Paper, Jim Odell

Jim reviewed the several additions to the Agent Technology Green Paper.

Areas that still need work:

Issues

Agent White Paper and Roadmap

We ran out of time to go over the current draft of the white paper and roadmap.  So we will do that in Denver.  We did discuss one of the three pre-RFP areas that we had planned to discuss:

Agent Discovery Service, Craig Thompson

At the Boston meeting, we decided to assign folks to dig deeper on proposed RFPs that seemed core to getting agent technolgy widely deployed.  These were agent identity, transport, and agent discovery and registration.  Thompson was charged with leading a discussion on the latter.  See Agent Discovery and Registration Notes  (internet/00-01-03).

Because of the overlap with the ECDTF Registration and Discovery RFP, we discussed the two together.  We generally think the ECDTF RFP will meet the needs of the agent community for the purpose of registering agents and finding them.  It will require that agents be represented by some standard means in XML Schema, at very least.  It was thought not to be too interesting to consider the discovery and registration service to be itself an agent.  Several ideas not in the RFP emerged:  compositional or inference objects, resource descriptions that span all of ontologies, the need for a family of query capabilities for text, properties, xql, prolog, or SL.  Possible need for plugins.  Also, we questioned if anyone would define "policy" the way it was described in the RFP, or know what was intended.

Agenda for Denver Meeting

... see Agent WG Denver agenda

Action Items