On Monitoring and Steering in Open MultiAgent Systems

1
On Monitoring and Steering in Open Multi-Agent
Systems
Takahiro Murata Naftaly Minsky Rutgers
University at SELMAS03 May 4, 2003
2
What is an Open Multi-Agent System(Community)

• A potentially large collection of autonomous
  agents that interact with each other, and with
  their environment.
• The community may be heterogeneous, consisting of
  people, and/or of software components designed
  and constructed independently of each other.
• Agents might join or leave the community
  dynamically, at run time.
• Community members might collaborate on a common
  goal, or compete with each other, or both.

3
Management in Open MASs

• An open MAS may have to be managed.
• The manager must be able to
• monitor the activities of its subordinates.
• steer the system, by controlling its
  subordinates.
• The question how can an heterogeneous,
  dynamically changing, community be managed
• particularly if the manger itself may be
  replaced, at run time.

4
An Example

• A department store deploys a team of agents, to
  procure merchandise.
• The team consists of a manager and a set of
  buyers.

5
An Example (contd)
6
An Example (contd)

• Question How to ensure that every buyer behaves
  as follows
• stays within its budget
• reports on its purchases to the manager
• obeys the managers commands to reduce its
  budget.

7
An Example (contd)

• But that is not all
• Question How to ensure the power of the manager
  is transferred properly, while the membership is
  changing?

8
What Needs to be Done?

• Establish a system-wide policy that stipulates
• conditions for joining and leaving the community
• budgetary limitations on purchasing
• the reporting of each PO to the current manager
• the managers authority to provide agents with
  budget, and to adjust this budget dynamically
• means for the manager to appoint a successor,
  providing the successor with the managerial
  powers, and ensuring that its identity is known
  to all team members.

9
What Does it Mean for a Policy to be Established?

• The policy needs to be specified explicitly.
• The policy needs to be enforced, so that it
  cannot be violated by any buggy or rogue agent,
  or by one who is oblivious of it.
• If the MAS is large, the enforcement mechanism
  needs to be decentralized, for scalability.
• This can be done via the Law Governed
  Interaction (LGI) mechanism.

10
Some Characteristics of LGI

• Openness self-organizable community, subject to
  its law with no inherent need for centralized
  administration.
• Regularities the assurance that every agent in
  the community conforms to the same law.
• Scalability completely decentralized policy
  enforcement.

11
Conclusion

• Management cannot be accomplished by
  role-assignment alone.
• It needs to be supported by a discipline that
  provides the manager with control over its
  subordinates.
• There are many types of such managerial
  disciplines, which can be established by LGI laws.

12
Conclusion (contd)

• Management is just one mode of coordination. We
  have experimented with many other modes,
  including
• collaborative work, such as decision making in
  distributed committee
• competitive interactions that call for
  sophisticated access control
• regulating the handling of alarms in a public
  domain
• and others . . .

13
Questions?
14
Distributed Law-Enforcement under LGI
15
Deployment of LGI
16
on the nature of laws under LGI

• Laws are defined locally, at each agent
• they deal explicitly only with local events at
  individual agents.
• Such as the sending or arrival of a message, the
  coming due of a previously imposed obligation,
  etc.
• the ruling of a law for an event e at agent x is
  a function of e, and of the local control state
  CSX of x.
• such a ruling can mandate only local operations
  at x.
• Yet, since law L governs every exchange of
  L-message between pairs of agents, it has a sway
  over the entire L-community, uniformly governed
  by L.