Washington DC
January 11-12, 1999
Co-chairs: Stephen
McConnell, James
Odell, Craig Thompson
Agent Working Group reports to ECDTF
and Internet PSIG
OMG Document ec/99-01-17
OMG Agent WG homepage: http://www.objs.com/agent
Precondition to becoming a SIG: get a green paper document - that captures agent architecture and relationships to current OMG services (which must be extended?) - that we can communicate to others.
Between now and the next meeting Thompson will draft a strawman SIG mission statement that is aggressive and aims at RFPs (just in case).
Odell to revise green paper based on inputs from Stout, Thompson, and others. Some suggestions for outline motivation section Introduction Purpose of Agent Group Working Defn of Agents and Important Agent Attributed Why Agents Matter What can OMA do to support agents What do we need to support agents Thompson and Odell to draft Agent Architecture RFI Thompson and McCabe to draft OMG-FIPA liaison Thompson to draft Agent SIG Mission Stmt in case we escalate to a SIG at PA meeting
Foils hand written. Recent paper on OMG server and sent to Agent WG last week.
MASIF deals with interoperability of mobile agent platforms, not mobile objects.
Mobile Objects have a lifecycle, and possible streamable and service provision interfaces. This architecture suggests a stateless bridge object which is an adapter from CORBA to implementation. Need unique IDs, security support, transaction support (optional). The ability to move objects from the US to the Caymen Islands for tax purposes is a legal EC issue, one of many motivations for mobility. Need security against hostile environments and agents.
CoABS is a 6 month old DARPA program focused on agents. One of the new ideas is the agent grid. DARPA’s problem domain is very large C4ISR problems. Agent technology is to componentize these problems and avoid stovepipes. Need to scale and dynamically hook together software agents (as opposed to intelligent agents). Two main fronts: semantic interoperability via declarative languages of communication and large scale multiagent systems (exploit parallelism, efficient use of resources, stable, uncertain environments). Tens of agents has been considered large in the past but with ALP we need millions of classes.
Broad vision of the grid: You are an infantry battalion commander who connects your PDA online. Your personal assistant agent connects to the grid. Your PDA tells your physical location and plans and resources. The grid adjusts to you. When you disconnect, it remembers you and gets ready for when you come back. There is a notion of goals and tasks. Might use Jini or other things as the plumbing. The grid is infrastructure to permit semantic interoperability. OAA and Retzina can talk. One agent or all might be grid aware. Grid lets you connect agent systems together. New components might be grid aware from the start. GITI and ISX are coordinating. Now: vision of the grid. Top down: what are the use cases of the grid. Minimal set of grid services. Initial services. Also what is out there, like OAA, Dartmouth D’Agents, … and looking at what they have. Retsina is looking at broker and trader agents. Implementation independent descriptions. Might be implemented on CORBA, Java, Jini, …
Build a platform to experiment with. Multiple implementations are desirable. Multiple implementations when security and fault tolerance are faced.
Simplifying assumptions - focus on functionality versus performance, ignore fault tolerance, … Scalability is paramount. Particular agent systems might employ control strategies. Also looking at minimal set of messages an agent on the Grid must support.
USE CASE approach
Story: a financial company only does trading with high bandwidth connection to its world-wide offices. A satellite driven trader office with one second round trip is too slow to make money.
The COS event service supports push and pull. The Notification service extends events with structured events, event filtering using filter objects, QoS configurability on per channel, per process and per message basis, event subscription information sharing between channels and clients, and optional event type repository (based on MOF). Can have customers that are event or notification service clients. You can have different administrators on a channel. Support for mapping to common events in SNMP, CMIP. Admin and Proxy objects can have filter objects. The constraint defn language is related to Trader
constraint defn language. Filtering can fail at the admin filter and still go on to normal filters. The can be and-ed or or-ed. Mapping filters are applied to QoS as opposed to content filters. The QoS filters affect priority and lifetime. Standard QoS properties for reliability, priority, order policy, and discard policy, starttime, stoptime, timeout. Channels keep track of who supplies and subscribes to event channels for various types - tells suppliers there are no customers until there is a consumer. If consumer registers and no one is sending, it tells you. Callbacks laced all the way through this. Provides either callbacks or polling, with blocking or nonblocking.
Document is on www.urep.com. Working for adoption at this meeting. Specification of XML externalization of UML and MOF. UML DTDs. Also MOF to XML Stream. See ad/98-10-05 and -06. Can transfer M3 (meta-meta model) down to instances. Used to publish metadata on the web. Or to transfer data across repositories.
This is a discussion for an hour. A paper by Odell with new additions by Jeff Bradshaw was circulated as a strawman during the meeting.
We started with a strawman list of technical aspects about agents. Then generalized to how are we proceeding and our goal. We need an agent architecture (landscape) and then prioritize. We also need to understand OMG Agent relationship to FIPA agent (as well as research communities like DARPA CoABS or EC Agentlink).
Lets not get too hung up on what is an agent. Instead lets locate kinds of functionality that we know we need.. One defn is that an agent is an encapsulated entity that has its own intentions and goals.
What will I do with agents and what wont I - billing system or logistics system.
Jim Odell and Craig Thompson will put together an RFI for Agent: technology landscape, new services needed, extensions needed (to UML). Key reason for RFI is credibility in the community and advertising. But it must be pointed.
A goal is to identify interfaces (to existing products).
A working group could discuss FIPA and OMG liaison: Francis McCabe, Jim Odell, Craig Thompson, Kate Stout,
Mappings of ACL or services or …
Key application areas of agents - possible workgroup for assembling scenarios and workgroups.
Relationship of RFI and green paper. RFI asks for inputs on architectural roadmap. Green paper is an initial response. And it contains a mapping to CORBA services.
There is no correlate of agent ACL to a current CORBA service. But there is Events, Asynch Comm, … Store and forward, multicast, broadcast, … transport mechanisms.
Francis talked about byte encoding in IIOP or IDL. Structured events.
Some question if IIOP is a good match with FIPA. Lots of discussion on ACL and whether it is part of the transport level or passed through by the transport level.
Other discussions occurred about aspects of the current green paper but not of the discussion was recorded in these minutes; rather Odell took notes on his copy of the green paper.