Agent Working Group, which jointly reports to Electronic Commerce Domain Task Force and Internet Platform SIG, met on Monday and Tuesday afternoon as scheduled. We heard presentations on agent mobility, the agent grid (DARPA Control of Agent Based Systems architecture), the COS Notification Service, and XMI: UML encoded in XML. Work items include contributions to a green paper on candidate agent RFPs, a plan to draft an RFI to develop a roadmap for agent related standards, and a mission statement for if we request to upgrade the Agent WG to a platform SIG at the Pennsylvania meeting. See separate Agent WG minutes.
Today different DoD logistics commands are not well connected and are based on often stale and not dynamic data reflecting C4ISR operational state. ALP is a 5-year DARPA project starting into its 3rd year that aims to radically change that.
Marshall distinguishes ALP’s general-purpose agent-based system architecture from the application domain of logistics. The application domain is global and solutions are federated.
There is a case-reasoning aspect or cascaded rule aspect that says to send a tank you must send people to operate it and so you need food, medical supplies, etc. The operational plan generates a log(istics) plan which consists of clusters. Most entities speak to a fixed number of other clusters so this does not turn into an N-squared problem. Clusters are agents that model some supply center. The ability to do good schedules is NP-complete from an optimization point of view but manageable through the use of naturally occurring hierarchies.
ALP is 100% Java built on Voyager. Most DoD logistics shops are NT shops or Unix shops, or …. The used Voyager to pass semantics and not just data. Individual shops might use IDL so that is a place where ALP talks to CORBA.
Clusters are being overlayed via a cost model.
Clusters are composed of plugins. Clusters are like agents. They are units of requirements. Some plugins break down a task into subtasks (expanders). Allocator plugins are like traders to associate needs with those who can execute them. Assessor plugins look at how well you are doing wrt the plan and actual. User interface plugs provide various visualizations. Data access plugins. JDBC plugins. Expert system generic plugins. Have not developed an agent framework for data access.
ALP uses a flexible object model. DoD has NSN catalog item numbers (unique ids for millions of kinds of things). This is too big (too many classes) for any compiler. So they use around 20 classes called prototypes. Then added to these are propertiesA 5 ton truck - with a hoist on it? Are there any objects that can hoist.
Clusters have encapsulation boundaries (isolation priciple) so one cluser cannot query into another cluster. Cluster A can allocate tasks but it is Cluster B’s business how you do execute them. So far they have only modeled a little of the logistics domain.
Clusters are a unit of planning. The focus this year is to provide feedback with reporting like, the best I can do is X with two days late cost. So you need human intervention.
Right now clusters are hard-wired with physical location and plugins. In the future, you want to dynamically allocate plugins and allow clusters to migrate along with mobile battilion so DBMSs move with the troops.
ALP has a lot in common with agent technology standards. It breaks a big problem into a society of individual clusters. No super agent knows everything. There is no thick book of log plan but rather a distributed view. No super coordinator. So it is hard to build the user interface to see the entire log plan. Testing an aggregate society is hard.
The ALP inter-cluster protocol is Assign/Reserve/Rescind. They need infrastructure on top to try out WHAT IFs. The need priority schemes., e.g., I’m a 3-star and you are 2-star. The cluster that rescinds a task rescinds its dependencies.
Agent ties might be strengthened so multiple clusters can cooperate, negotiate, … richer protocols.