FOUNDATION FOR INTELLIGENT PHYSICAL AGENTS
Domains and Policy Work Plan
FIPA Architecture TC
Date of this status:
FAB comments (see end)
This work plan is intended to address the issues associated with controlling agents within large multi-agent systems.
Problem Statement: Developers and users of multi-agent systems often wish to place strong constraints on the behavior of agents within agent environments. This includes being able to apply and enforce these constraints and policies across distributed agents and systems. The general focus of this work plan is the application and management of policies and constraints on agents and collections of agents, not the detailed management of agent lifecycle, and areas currently addressed by FIPA agent management specifications.
Objective: To develop mechanisms that:
· Describe desired behaviours declaratively,
· Allow agents to reasons about the constraints on their behaviour,
· Allow humans to reason about the constraints on the agents and systems. (intelligibility),
· Enforce behaviour for agents on multiple hosts,
· Permit changing constraints without changing code,
· Support explicit constraints,
· Separate policy from mechanisms,
· Describe the constraints,
· Apply behaviour to groups of agents (as small as one),
· Common point of administration.
· Policy language(s) and notations
· Architectural abstractions for policy distribution and enforcement
· Domains for grouping agents
· Formal verification methods for testing conformance
We will approach this from two perspectives:
· The expression of policy and constraints and way in which we reason and manipulate them.
· The domain of discourse – what are the proper subjects of the policies and constraints (set of behaviors, characteristic attributes of agents and agents systems about which we wish to create policies)
We will need to provide mechanisms for describing what policies can be applied to certain agents, and in certain environments. (e.g. what controls are available, and how to report inability/conflict when applying policies)
The range of possible mechanisms for enforcing policy mechanisms we contemplate supporting extend from reputation and social sanctions to complete withdrawal of supporting services for a non-conforming agent. We intend to explore use cases in support of both pre-condition constraints and post-condition enforcement.
Specifications Generated: The goal of this work plan is to generate architectural abstractions in support of control-ling the behavior of agents in multi-agent systems. This will include constraint languages for describing the behavior of the agents and extensions to the abstract architecture to enforce these behaviors.
Plan for Work: Study problem, solutions in existing systems, including KAoS, Magenta, April, ??? ex-tract design patterns based on these systems. Develop requirements for specific types of constraints and policies. (e.g. All agents which send messages outside their local environment must encrypt these messages). Examine domain and related structuring techniques for relevant design patterns.
· Create a contract model in terms of FIPA communicative acts and predicate calculus
· Create an ontology of the contract framework (constraint/policy framework)
· Create specific ontologies for specific domains
· Create set of architectural elements to distribute and enforce contracts.
· Provide guidance on writing reasonable and intelligible policies
· Create a basic domain model for FIPA systems, including mechanisms to define and apply policies to groups of agents. .
· Provide test cases for testing environments which implement these extensions.
· Provide libraries of basic policies.
FIPA Rationale: FIPA provides a framework for creating inter-operating agent systems. At the same time, most existing agent systems which have been deployed, have found compelling reasons to define, apply and manage policies for their systems. If we do not provide these mechanisms within the FIPA standard, policy mechanisms will only be applicable within the individual sub-systems created by FIPA members, rather than across multiple FIPA systems.
· 2000/04 Create work plan, issue call for information.
· 2000/11 First draft of ontologies and contract frameworks, proposed set of architectural elements.
· 2001/01 Preliminary document.
Based on progress during the next year, we will define additional milestones for this work effort. Success criteria will be based upon the successful implementation of these ex-tensions within the FIPA community.
· [FIPA00001] FIPA Abstract Architecture Specification
· [FIPA00061] FIPA Agent Communication Language Parameters Specification
· [FIPA00037] FIPA Communicative Act Library Specification
· [FIPA00025] FIPA Interaction Protocol Library Specification
· [FIAP00007] FIPA Content Languages Specification
· Frank McCabe, Fujitsu Laboratories of America
· David W. Levine, IBM
· David Evans, Imperial College/Simplex Technology
· Jeff Bradshaw, Institute for Human and Machine Cognition/UWF
· Jim Odell, James Odell Associates
· Geoff Arnold, Sun Microsystems
· Kiyosho Kogure, NTT
This work plan has been approved and assigned to the Architecture Technical Committee.