Interaction Monitoring and
Termination Detection for
Agent Societies
Preliminary Work
T. Motshegwa, M. Schroeder
Agents@City Group ,
City University, London, UK
1
Overview: Structure of the talk
Motivation
–
–
–
Problems
Goals
Scenario (s)
Distributed Termination Detection Research
Agents, Protocols & Termination Detection
Definitions
Framework, Algorithm and Implementation
Future Work & Summary
Working Assumptions
Computing Pespective: Agents are autonomous,
asynchronous , distributed and possibly mobile
processes.
Behaviour specification through protocols.
Agents communicate in an ACL.
Agents merely observed from the outside.
–
–
Observed Communication
Elements of the behavior specification made available
MAS: Modular Distributed Systems.
–
–
Decentralised data
Agents have incomplete information or capabilities
Scenario
Client/Buyer (Bids)
Server/Seller (Asks)
Trader
Trader
Securities Market
Electricity Market
Market
place
(CDA)
Trader
Trader
Computational
Trader
Resource Market
Service Market
Trader
Introduction: Problem
Interactions and termination detection in agent societies
E.G Deciding when the negotiation process is completely
terminated.
Also: determining a global view from partial local views
Two Aspects (approaches) to the problem.
– Formal Verification of behavior at compile time.
– Monitoring and Control at run-time.
In general Verification is difficult and not satisfactorily
solved for deployed systems
Also consider Monitoring and Control to complement
formal verification research.
Goal: Interaction Monitor
Development of a MAS Interaction Monitor.
Agent Designer
–
–
–
–
–
Visibility of interactions
Stage of task execution or Goal achievement
Current Levels of resources
Performance of strategies.
Control Tasks e.g returning, terminating , updating parameters
Society Manager
–
–
–
–
Automatic resource management and scheduling mechanisms
Management of registries, through timely garbage collection of defunct
agents.
Agent Pooling and facilitation of optimal execution.
Enforcing market rules and observing deviant behaviour.
Distributed Termination Detection
Research Overview
Classical Problem of Distributed Control.
Of Practical , Algorithmical, theorical and methodical
importance.
Communicating process model: Termination implies ALL
processes idle and NO computation is possible.
Example of a STABLE property (Property of the global
state that if ever becomes true , will remain true forever).
Related to deadlock detection, GC, Snapshot computation,
distributed Consensus etc.
Determining termination is a non-problem in sequential
world
Cont..
Very Difficult to check in a distributed whether all
processes are passive.
Difficult to achieve efficiently, it’s a GLOBAL
state detection problem
A global state of DS consists of:
–
–
State of each process / object
State of channels
Difficult because of synchronisation problems (No
shared memory , shared clocks etc.)
Algorithms: Tracing and Probe-based
1) Tracing: ‘For Diffusing Computations’ (Dijkstra)
–
–
–
–
–
–
–
–
–
–
‘Environment’: Supplies single input to an arbitrary process
(source)
Arrival of input action can trigger local and output actions at a
process.
Computation Diffuses in a system due to stimulus.
Maintenance of a Spanning tree.
Messages ACKed and deficit count maintained.
Termination: NO process is enabled to perform local action and NO
messages in channels and all messages ACKed.
Quiescent nodes with all messages ACKed delete themselves.
Tree grows and shrinks.
If deleting node == source : TERMINATION
EG. Dijkstra Scholten’s Algorithm.
Algorithms: Taxonomy
2) Probe Algorithms
– Characterized by a ‘wave’ execution
– Waves of message passing collect information
about global state
– Special node ‘Controller/Initiator’ envisaged.
– Controller co-ordinates detection
– Exploit feature of network topology, e.g token
Ring.
– E.G Dijkstra, Feijen Van Gasteren algorithm
Agents, Protocols, Termination.
Agents are higher level entities cf Objects.
– Behaviour Specification in Protocols
– Execution of protocols and complex interactions.
Standard Algorithms cannot be used directly.
–
–
Algorithms abstract away from PURPOSE and OPERATION of
Computation
Simple model of underlying computation.
Agents may have additional ‘knowledge’ they can
reveal to aid termination detection.
Protocols.
Regulate Interactions.
Viewed as Specifications of these interactions.
We model protocols as labelled state transition
system with a set of terminal state (S,
,L,T)
EG. CNP ,S ={1,2..,8}; T={5,6,7,8}; L={cfp,
propose..} and
A:cfp
B:Propose
A:Accept_proposal B:inform
=
1
2
3
4
B:refuse
A:reject_proposal
B:not_understood
6
7
5
B:failure
A:cancel
8
Termination Detection For Protocols.
Definitions
Termination Path, (TP)
Unique Termination Path, (UTP)
Shortest Unique Termination Paths, Observables
Example.
1
c
p=(1,2,5) valid TP with labels (b,c)
p=(2,5) is TP but not UTP
BUT p=(1,2,5) is UTP
NO other path with (b,c)
2
a
1
a
b
3
2
4
d
e
c
a
NO UTP
5
Termination Detection For Protocols.
Shortest Unique Termination Paths
Set of Observable O = {b,c,d,e}
NB: a is not an element because
SUTP in (1,3,5) is (3,5) labelled (d)
1
b
3
2
4
d
Minimal information (Sub-protocol)
Needed is SUTP of Agents
e
c
5
Proposed Framework
Deliberative detection framework similar to probe-based
TD algorithms.
Monitor queries agents for information about protocols
and their execution.
Querying mechanism based on polling.
Monitor can reconstruct global view or deduce termination
of protocols based on information.
ASSUMPTIONS
–
–
–
–
Protocols used in MAS are KNOWN.
Agents have incentives to communicate accurate information
Monitor can forcefully terminate or suspend agents whose behaviour
deviate from rules of protocol.
In principle monitor can listen to all communication BUT periodic
polling is used to minimise communication overheads.
Algorithm.
Given: TP be the set of Shortest Unique Termination
Paths, SUTP, with Observables, O.
Data Structures: For every p = (s1,…,sn)  TP there is
a state si called the current execution state of p.
Init: For all p = (s1,…,sn)  TP initialise its current
execution state to s1.
Algorithm cont..
Body: While not TERMINATED do
–
–
–
–
–
Let l be a message be sent by an agent
IF l O then for p = (s1,…,sn) TP set p’s current execution
state to s1
ELSE: For all p = (s1,…,sn) TP with current execution state
si DO
* IF si l si+1 then p’s execution state becomes si +1
and if si+1 = sn then set terminated to true.
IF si l si+1 then set p’s current execution state to s1
or to s2 in case that s1 l s2
Initial Prototype Implementation.
Design Options , Broadcasting Vs Network topology
e.g Token Ring.
Initial Prototype implementation based on
Broadcasting.
In theory ,worst case scenario in message complexity
Polling used to reduce message traffic.
Polling frequency , parameter for experimentation
Suitable for less dynamic applications.
Uses JADE agent execution model.
Related Work
Quiescence detection in multiagent multi-issue
negotiation [Wellman].
Uses the tracing Dikstra Scholten Algorithm as a basis of
detection protocol.
Protocol operates as a layer on top of an underlying
mediated negotiation protocol.
Overall agent’s behavior is a composition of basic
negotiation behavior with state transition diagram
representing algorithm.
Issues & Future work
If Detection mechanism assumes correct participation
by agents
–
–
–
Unilateral (maybe strategic) deviation is a concern.
Monitor knows protocol specification and enforces rules of participation.
Consider incentives for agents to participate.
Monitor as a bottleneck and single point of failure.
–
–
–
How to decentralise?
General problem for most services eg Directory, Naming Services
There exist Solutions e.g hierarchical setup in DNS.
Experiments
–
–
–
Detection Delays.
Scalability.
Information Revealed by Agents & Market Rules.
Summary
First Step in realising a Framework for studying
determination of a global state such as termination in
agent societies.
Linking research is DS and MAS.
Suggested Definitions and minimal additional
information that agents need to reveal to aid the
detection process (SUTP).
Identifying Outstanding issues and future experimental
work.
Generate Discussion.