Monitoring Agents

1
Monitoring Agents

• Ying Qian Zhang/ Jürgen Dix
• The University of Manchester
• Thomas Eiter/Michael Fink/Alex Polleres
• Technical University of Vienna

2
Outline

• Motivation
• Monitoring Agents in Multi-Agent System
• Proposed idea
• IMPACT Agents
• Model an IMPACT Multi-agent System
• K-Planning Language
• Build a Monitoring Agent
• exploit DLVK
• Implementation and Demo
• Monitoring a Gofish Post Office MAS
• Perspective and Discussion

3
Motivation

• How to model/specify a multi-agent system in
  order to reach the desired goal?
• How to guarantee the MAS is working in a good
  state? How to verify it?
• How to make a MAS intelligent so that it can
  detect/diagnose/recover from the failure?
• The performance of the communication of agents is
  critical to a Multi-agent System.
• In a MAS, convert the modelling problem to a
  planning problem.
• Use the planning system to specify/monitor the
  agents collaboration for a particular
  application.

4
Monitoring Agents Idea

• Goal
• Exploit planning systems to model and monitor the
  collaboration of agents in a MAS
• How?
• for a particular application
• Actions in the plan Messages sent between
  agents
• Check the consistency of actions ?
• Monitor the collaboration of agents

5
Monitoring Agents Idea

• Goal model/monitor a good collaboration of
  agents in a MAS

Action Language K
DLVK
Multi-Agent System
A Monitor Agent
Planning Problem
Is MAS in good states?

• The idea of monitoring is independent of
• A MAS ( IMPACT ) or
• Planning language ( language K)

6
IMPACT

• IMPACT Interactive Maryland Platform for
  Agents Collaborating Together
• (Subrahmanian/Bonatti/Dix/Eiter/Kraus/Özcan/Ross)
  
• how can it help?
• Access to distributed, heterogeneous data sources
• Seamless interoperability between different
  software capabilities
• Ability to coordinate multiple agents

7
IMPACT Agents Data Access

• Code Call df( arg_1,,arg_k )
• Code Call Atom in( X,df(arg_1,,arg_k) )

Execute function f defined in agent d on the
specified arguments. Returns a set of objects.

• Kplan--gtKplancalcKplan ( D\bin\DLV\Kplan\gofi
  sh.plan, D\bin\DLV\Kplan\gofish.bk,
• "-FPsec", 14, "-silent" )

Succeeds if X is in the set of objects returned

• in ( plan, Kplan--gtKplancalcKplan
  ("D\bin\DLV\Kplan\gofish.plan",
  "D\bin\DLV\Kplan\gofish.bk",
• "-FPsec", 14, "-silent" ) )

8
IMPACT Agents

• An IMPACT agent consists of
• a set of data types
• a set of API functions implemented in any
  language manipulating those types
• a set of actions implemented in any language
• a notion of concurrency
• a set of action constraints
• a set of integrity constraints
• an agent program

9
Action Language K for formulating a
planning problem

• The language K (Eiter et al.2000)
• a logic-based planning language, syntactically
  similar to the action language C, but
  semantically closer to answer set programming
• is well-suited for representing incomplete
  knowledge
• with parallel and sequential actions
• constructs optimistic and secure plans

10
DLVK Planning exploited by a particular
monitor agent

• DLV (http//www.dbai.tuwien.ac.at/proj/dlv)
• A deductive database system, based on
  disjunctive logic programming, which offers
  front-ends to several advanced KR formalisms
• (http//www.dbai.tuwien.ac.at/proj/dlv/K)
• A front-end of DLV for K-planning language
• K-Planning frontend option
• -FPsec generate secure plans
• -FPopt generate optimistic plans

11
Monitoring Agents

• Just starts
• Application specific
• A small application is implemented in IMPACT
• Monitoring Gofish Post Office MAS, which
  simulates the package delivery service

12
Monitoring Gofish Package Delivery

• Gofish Post Office
• Allows citizens to send certain kinds of
  packages to other citizens in Gofish. Packages
  are dropped, shipped, and delivered.
• Gofish plan problem formulation using Language K
  (A.Polleres et al.)

13
Monitoring Gofish Package Delivery

• Background description

• Gofish Plan description

gofish.plan fluents undelivered.
packageAt(Pid, L) requires package(Pid,PType,Wt,
...,DelivTime), location(L). delivered(Pid)
requires package( ) actions DropOff(Pid,
PType,) requires package(Pid ...) always
caused undelivered if not stats_sent(Pid).
inertial packageAt(Pid,L). inertial
added(Pid). noConcurrency. goal
not undelivered?
gofish.bk ev(dropoff). ev(distcenter).
ev(delivery). ev(truck). location(DropOff).locat
ion(ontheway). location(DistCenter).
location(Truck). location(Dest). package(pidVa
r,pTypeiVar, wtVar, volVar, lSenderVol,
fSenderVol, lRecipVar, fRecipVar,
recipTelVar, recipEmailVar, origZipVar,
destZipVar, destStreetVar, destNumVar,
priorityVar, costVar, dropTimeVar,
delivTimeVar).
PLAN DropOff(Pid,) add_package()
send_ziptozip() DistCenter()
send_centertohouse() get_recipient_info()
recipient_info() Truck() send_trucktohouse()
get_recipient_info() Delivery()
set_delivery_time() statistic_info().
command line DLV gofish.plan gofish.bk FPsec
planlength14
14
Monitoring Gofish MAS Design

• Gofish Multi-Agent System
• Developed to simulate the package delivery
  service of Gofish Post Office (TU students)
• Package Agent Notification Agent Zip Monitor
  Agent Event Manager Agent Dispatcher Agent
  Package Dropper Agent
• Agents Asynchronous interaction via message
  exchange
• Gofish message protocol

15
A Monitoring Agent how to work

• Some functions
• obtain a set of plans from DLVK
• at regular interval, read messages sent between
  agents in Gofish MAS from the actions log file
• compare messages with actions in the plan, if
  inconsistent, give out failure message
  otherwise, continue monitoring

16
A Monitoring Agent Implementation

• Software Code
• new connection
• data type definitions
• API function calls
• Agent Actions
• actions that can be executed by agent
• Agent Program
• a set of rules

• K-planning
• exploits DLV K-planning frontend
• Plan type
• Code-calls
• get plans from DLVK
• read log file of Gofish MAS
• check the state of Gofish MAS
• consider a plan
• reduce a plan
• failure message
• success message

17

• Demo
• Monitoring Gofish Multi-agent System

18
Future Work Discussion

• Monitoring
• The monitoring agent should monitor several tasks
  in parallel
• We only detect/report the failure of the MAS
• Recovery/re-schedule may be considered
• A debugging tool to the MAS
• Plan Types
• We used secure plans in our first application
• Consider other types of plan, i.e. optimistic plan

19
Future Work Discussion

• Collaboration of Agents in MAS
• The monitoring agent only watches the
  communication between agents
• Other actions executed inside agent(s) may be
  monitored in the future
• Could it be a feasible approach towards
  verification of a MAS?

20
Thanks