Agent Framework Technology

1
Agent Framework Technology

• An overview of agent frameworks for use in
  GasTurbnLab

By Matthias Wollnik
2
Overview

• 1. Agent definition
• 2. Agent types
• 3. GasTurbnLab agents
• 4. Agent frameworks
• 5.Conclusion

3
What is an Agent?
1) Process 2) Autonomous 3) Very specialized 4)
Originally meant for AI work 5) Recently used for
collaboration and internet searches
4
What can an agent be?
1) Mobile 2) Can communicate with other agents 3)
Intelligent 4) Distributed 5) Extensible
5
Agent Types
1) Mobile Agents 2) Collaboration 3) Search
Agents 4) Cognitive Simulation 5) Miscellaneous
6
Mobile Agents Theory

• Need to get and process information from many
  machines
• Can be a problem when the connection is not good
• Solution Dont communicate over uncertain
  connections
• Move the agent and communicate locally

7
Mobile Agent Structure
Compiled Agents
Interpreted Agent 1
...
Agent Virtual Machine
Host Operation Systems
8
Collaborative Agents
Data-Space Sites
Data-Processing Sites

• Workstations provide data
• Can change the data that everyone sees

9
Agent use in Collaboration

• communication between workstation and data-space
• communication between data-space machines
• communication between data-processing machines
• The agents allowing the data to be seen from each
  workstation require that they get their
  information from the data-space which must
  distribute it using agents.

10
Search Agents

• Inspired by the vastness of the Internet
• Move from machine to machine searching for data
• Commonly implemented as locally running agents
• Online search engines are biggest user
• may be used to prefetch web pages

11
Cognitive Agents

• Commonly used in the AI community
• Often set up to simulate intelligence models
• Not usually possible to handle external access
• Use proprietary languages (or prolog/lisp) to
• describe models

12
Miscellaneous
There are other types that are meant for
specific purposes like gathering information out
of your local PIM. Also general frameworks that
can emulate two or more of the previous ones fall
in this category.
13
Agent Summary
14
GasTurbnLab Agents
Basic Requirements
1) Distributed agents 2) Communication between
agents 3) Ability to run Kiva, Ale3D, and
//ELLPACK code 4) Ability to build mediators with
either internal language or compiled code
15
GasTurbnLab Agents
Possible Extensions
1) Agent intelligence 2) Dynamically assign
agents to machines 3) Split up problems if new
agents become available 4) Add machines to
resource pool at run time
16
Agent Frameworks
Soar

• Started in 1983 at CMU
• A workshop is held every 8-10 months
• Main institutions involved
• CMU
• Ohio State
• UMich
• Univ of South California
• multiple sites in Europe
• In use at over 100 sites around the world

17
Agent Frameworks
Soar

• Mainly used for cognitive research
• has been used to simulate learning
• psychological models
• Supports the creation of autonomous processes
• allows very extensive intelligence modelling
• allows extension building via Tcl
• communication between agents is done by I/O
  routines
• does not support network communication
  inheritently

18
Agent Frameworks
Soar Sample notation
Sp applymove-blockremove-old-ontop
(state ltstate1gt operator ltop1gt
ontop ltontopgt) (ltop1gt name
move-block moving-block
ltblock1gt destination-block
ltblock2gt) (ltontopgt top-block ltblock1gt
bottom-thing ltgt ltblock2gt
ltblock3gt ) ----gt (ltstate1gt ontop
ltontopgt -)
19
Agent Frameworks
Grashopper Agent Platform

• Made by the German company IKV
• (Informations- und Kommunikationstechnologie)
• Grasshopper is a mobile agent framework
• Compliant with OMGs MASIF standard
• based upon CORBA

Compiled Agents
Interpreted Agent 1
...
Agent Virtual Machine
Host Operation Systems
20
Agent Frameworks
Grashopper Agent Platform

• Written in Java using ORB
• can be extended easily using ORB applications
• supports regions which allow sub-partitioning
  of network
• supports compiled (service) agents and Java
  agents
• supplies tools to get information on running
  agents on network
• region registry is used to find other agents
• Commercially supported and available for 3000
  for binary
• developers license. Source code license costs
  110000

21
Agent Frameworks
Ara-Platform for Mobile Agents

• Made by the University of Kaiserslauern, Germany
• Supported by 6 people under Nehmer
• implemented in C
• Supports compiled languages as dynamically loaded
  libraries
• First publication on Ara is dated 1997
• Web page very out of date
• Stationary agents are used as services that
  mobile agents
• come to interface with

22
Agent Frameworks
Ara-Platform for Mobile Agents
Sample Agent in C int main(int argc, char
argv) char homeStringARA_PLACENAME_SP
ACE Ara_PlaceName home Ara_Here(),
destination Ara_PlaceNameCreate("thor")
Ara_Go(ARA_ID_NONE, destination,
ARA_ALLOWANCE_NONE,
ARA_ALLOWANCE_NONE) Ara_PlaceNamePrint(home,
homeString) printf("Hello at thor, I've come
from s!\n", homeString) Ara_PlaceNameDelete(
home) Ara_PlaceNameDelete(destination)
Ara_Exit(0) / Never reached /
return -1
23
Agent Frameworks
Ara-Platform for Mobile Agents
Sample Agent in Tcl ara_agent set home
ara_here ara_go thor puts "Hello at
thor, I've come from home!" ara_exit
24
Agent Frameworks
Collaborative Computing for Nat. Sci. Research

• Built at Emory University
• Supported by 17 people under Sunderam
• Collaborative environment
• written in C/C for Solaris and IRIX
• Supports compiled languages as stationary
  services
• Will upload programs to machines designated to
  run them
• Uploadable agents written in either Perl or Java
• Basic message passing interface

25
Agent Frameworks
Bond Distributed Object System

• Built at Purdue University
• Originally built for Computational Biology
• written in Java
• Supports directly callable external programs
• Active development by three grad Students with
• supervision by one faculty member
• Cannot easily be classified in any one category
  of Agents
• Message passing is done via KQML over infospheres

26
Agent Frameworks
Bond Distributed Object System
Each agent is implemented as a finite state
machine in either Java or a proprietary
format public Class PDESolverAgent extends
bondAgent public PDESolverAgent()
bondStrategyGroup bsu1 new bondUtilStrategyGroup
(fancy1) fsm.addState(new
bondFSMState(name,bsu1.getStrategy(name)))
fsm.addTransition(name1,name2,condition)
.
27
Agent Frameworks
Bond Distributed Object System
begin blueprint agent PDEMediatorAgent add state
AgentStarted with strategy Util.WaitAndSucceedde
fault internal transitions from
AgentStarted to DataStaging on success from
AgentStarted to Error on failure
28
Agent Frameworks
Bond Distributed Object System
The system has the unique advantage that we
already have a sample implementation for
SciAgents done. However, there have been serious
problems with stability in the framework.
29
Agent Frameworks
Open Agent Architecture

• OAA is being developed at SRI on IRD funding
• by Cheyer and Moran.
• Meant to be a general framework to produce a
  variety
• of agents including mobile, distributed, and
  learning.
• Exists as set of libraries for Prolog, C,
  Common Lisp,
• Java, Delphi, and Visual Basic
• Communication via InterAgent Communication
  Language
• Very little documentation available to the public
  at this time
• Release date of April 99
• Free for non-commercial use

30
Agent Frameworks
Open Agent Architecture
System will use a Facilitator design. The
Facilitator will receive requests and make
decisions as to which agent the work is assigned
to. This allows the Facilitator to split complex
problems into pieces as more agents become
available.
31
Agent Frameworks Summary
32
Agent Frameworks Summary
Addendum ObjectSpace Voyager
ObjectSpace has built a free mobile agent
framework named Voyager. They receive revenue by
building programs on top of Voyager. Voyager is
very similar to Grasshopper in that it is
implemented in Java using CORBA. However, it
does not support external programs. Has a very
nice multicasting setup and has RMI support.
33
Conclusion

• For our purposes, we can use either a distributed
  or a mobile
• paradigm.
• The approach taken by CCF is not advisable for us
  since it
• results in message-passing be a major focus for
  us.
• We would have to modify all solvers to use the
  functions
• or write wrappers.
• True also of OOA, but added advantage would be
  Facilitator
• agent
• Soar is not a viable option since it allows no
  network access
• and does not allow interaction with legacy code

34
Conclusion

• Bond has distinct advantage since SciAgent has
  been ported
• Unreliable support
• Ara would be a reasonable system, however getting
  more
• current data is important
• Grasshopper may be the best solution if money is
  available
• since it is backed by a company

35
Conclusion

• If mobile Agents are acceptable Grasshopper
  should be
• considered as most useful agent framework
• Otherwise the OAA framework will most likely be
  best
• If the limited support is acceptable using Bond
  for the
• first implementation would also be acceptable.