V.S. Subrahmanian

1
IMPACT Interactive Maryland Platform for Agents
Collaborating Together

• V.S. Subrahmanian
• Department of Computer Science
• University of Maryland
• vs_at_cs.umd.edu

Part funding from Army Research Lab, DARPA,
University of Maryland, and some non-government
sources. Research initiated by Army Research
Office and Army Research Lab
1
2
IMPACT Technical Overview

• IMPACT Agent Architecture
• What is an IMPACT agent?
• What is required of a platform supporting
  multiagent applications?
• Key Scientific Contributions
• Agent development languages
• Temporal and uncertain agent reasoning
• Agent security
• IMPACT Agent Development Environment
• Key Applications
• ARL Combat Information Processor
• Army Logistics
• Agent based security
• Distributed Simulations
• Planned work for Years 3 and 4

3
IMPACT Agent Architecture
Agent Program
Actions
Concurrcy
Function Calls
MSG Box
Legacy/Special Data Structures
Action Constraints
Integrity Constraints
Other agents
(distributed) data - agent state
NETWORK
4
Summarized Agent Definitions

• Many definitions of agents have been proposed. We
  outline three representative examples - FIPAs,
  DARPAs and Shohams.
• FIPA Proposed a list of basic agent
  capabilities, incl.
• speech, visual, actuator I/O primitives
• syntactic and semantical info fusion
• messaging languages, interaction protocols, etc.
• Darpa Co-ABS
• agents act autonomously to accomplish objectives
• agents adapt to their environment.
• agent cooperate to achieve common goals.

5
Agent Definitions

• Shohams definition- an agent is a program
  supporting
• Ongoing execution
• autonomy
• adaptiveness
• intelligence
• agent awareness
• mobility
• anthropomorphism
• reactivity
• evaluating courses of action
• communication
• planning
• negotiation

All the above definitions are behavioral
definitions. IMPACT provides a structural
definition of an agent, together with
formal models of the desired behaviors specified
by FIPA,Co-ABS and Shoham.
6
IMPACT Agents Data Access

• Code Call df(arg1,,argk). Execute function f
  defined in data structure d on the specified
  arguments. Return a set of objects.
• Oracleselect(depots.rel,item,,combat boots)
• Routeplan(map1,20,20,50,43).
• Code Call Atom in(X,df(arg1,,argk)). Succeeds
  if X is in the set of objects returned.
• In(X, Oracleselect(depots.rel,item,,combat
  boots)). Find all tuples with item field equal to
  combat boots
• in(X, Routeplan(map1,20,20,50,43)). Find all
  routes returned by route planner between points
  (20,20) and (50,43) w.r.t. map1.

7
Code Call Conditions

• A conjunction of
• code call atoms and
• comparison atoms.
• In(X, Oracleselect(depots.rel,item,,combat
  boots)) X.qty gt 1000 in(R,routeplan(map,99,
  25,X.xc,X.yc)).
• The above says find X,R such that
• X is a tuple in an Oracle depot relation
  specifying a depot with over 1000 combat boots,
  and
• Y is a route from the location of that depot to a
  location (99,25) where the entity requesting the
  combat boots is located.

8
Agent Integrity Constraints

• An agent integrity constraint is a condition that
  the agents state must always satisfy.
• An agent integrity constraint has the form
• ltcode call conditiongt gt ltcode call atomgt
• If the agent state satisfies the code call
  condition, then it must also satisfy the code
  call atom.
• All created routes must be over 90 safe.
• In(R,routeplan(Map,X1,Y1,X2,Y2)) gt
  In(S,routesafe(R)) S gt 0.9
• Everybody makes less money than their boss.
• In(T1,oracleall(emp)) in(T2,oracleselect(em
  p,name,,T1.boss)) gt T1.sal lt T2.sal

9
Agent Actions

• An action consists of
• action name a(a1,...,an)
• a precondition code call cond.
• An add list code call cond.
• A delete list code call cond.
• An execution algorithm
• Unlike AI, actions are implemented via imperative
  code.
• Preconditions/Add/Delete lists are used by
  IMPACT to assess effects of the action, not to
  implement the actions.
• Each agent has a set of associated actions.

• Examples of actions
• execute request
• modify request and execute
• send message(s)
• clone agent
• move to remote site
• do-nothing
• post web page
• create file
• create bar graph
• construct route
• update database
• ETC..

10
Agent Concurrency Notion

• Takes several actions as input.
• Returns a single merged action as output.
• IMPACT has proposed 3 ways of concurrent action
  execution.
• They vary in underlying intuitions and
  application use, but also
• one is polynomial (efficient !)
• one is NP complete
• one is co-NP complete.
• Lesson Choosing sophisticated ways of merging
  actions may be expensive.
• IMPACT Agent Development Environment supports a
  library (under development) for the agent
  developer to choose concurren execution methods.

A c t I o n s
A c t I o n
Notion of concurrency
Agent state
11
Agent action constraints

• Action constraints specify conditions under which
  the agent cannot concurrently execute certain
  actions.
• They have the form
• a1,,an lt ltcode call conditiongt
• This says that if the code call condition is true
  w.r.t. the agents current state, then actions
  a1,,an cannot be executed concurrently.
• Examples Moves in two different directions
  cannot be concurrently executed.
• move(Dir1),move(Dir2) lt Dir1 ltgt Dir2.
• Orders must stay within the budget.
• order(Item1,Cost1),order(Item2,Cost2) lt
  in(Amt,budgetavail()) Amt lt Cost1 Cost2.

12
Agent Program

• Set of rules of the form
• Op a(arg1,,argn) lt ltcode call conditiongt
• /- Op1
  a1(ltargsgt) /- Opn an(ltargsgt).
• Op is a deontic modality and is either
• P - permitted
• O - obligatory
• Do - do
• F - forbidden
• W - obligation is waived.
• If the code call condition is true and the
  deontic modalities in the rule body are true,
  then Op a(arg1,,argn) is
  true.

13
Example agent program rules

• Do send-warning(E,M) lt in(Loc,entitycurloc(E))
  
• 
  in(E,entityhostiles())
• 
  in(Loc,entitycurloc(E))
• 
  in(D,geodist(Loc,Loc))
• lt(D,25)
  
• 
  (M,filecreate(E,Loc,D)).
• Do log(E,M,T) lt Do send-warning(E,M)
• in(T,clocknow())
• F logging.
• Send a warning to entity E if there is a hostile
  entity within 25 distance units of it.
• If a warning is sent and logging is not
  forbidden, then send a log message now.

14
What is an agent?

• 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
• IMPACT agents can provably support the criteria
  listed in the behavioral definition of an agent
  given by Shoham, FIPA, IBM, etc.

15
Agentization Procedure

• To convert a program to an agent, do the
  following
• describe types manipulated by program
• describe I/O types of API function calls
• define agents integrity constraints
• select/define actions that can be executed by
  agent
• select/define concurrency notion
• define agents action constraints
• define agents agent program
• In addition, one may
• specify yellow pages info and register agent with
  IMPACT Server
• specify connect information

16
Advantages of IMPACT Agents, I

• Rich mathematical foundation. Demonstrated by two
  papers in AI journal, one accepted paper for J.
  of Logic Programming, and accepted book on IMPACT
  (MIT Press, spring 2000), and several submitted
  journal papers.
• Can build on legacy data/code and specialized
  data structures. Demonstrated today by
  applications on Army Logistics data, Army JANUS
  simulation data, ARL CIP servers, Oracle data.
• Dynamic - can couple arbitrary actions to changes
  in agent environment. Demonstrated today by
  coupling of send-mail, create Web pages, file
  synthesis actions to changes in agent state when
  new messages are received.
• Open - can interact with other agent platforms.
  Demonstrated today by IMPACT connectivity to
  arbitrary IBM Aglets. Specifically, shows IMPACT
  security agents interacting with security agents
  built by Lockheed Sanders Mike Gaughan (ATIRP
  Consortium) and ARLs LTC. Paul Walczak using the
  IBM Aglet agent paradigm.

17
Advantages of IMPACT Agents, II

• Security - IMPACT agents can be used to make
  other applications more secure. Demonstrated
  today via ARL-ATIRP-IMPACT security agents.
• Intelligence IMPACT agents are capable of
• collaborating with one another as well as agents
  built in other agent platforms. Demonstrated
  today via bookstore procurement demonstration.
• creating sophisticated plans. Demonstrated today
  via noncombatant evacuation ops demonstration.
• reasoning about time and uncertainty.
  Demonstrated via research papers and Prof. Dixs
  lecture tomorrow.
• making decisions based on agent objectives
  (expressed via objective functions). Demonstrated
  via research papers and Prof. Eiters lecture
  today.

18
Advantages of IMPACT Agents, III

• IMPACT agents can be
• Autonomous. Demonstrated today via logistics
  agents.
• Reactive. Demonstrated today via bookstore
  agents.
• Mobile. Demonstrated today by IMPACT security
  agents.
• IMPACT agents support heterogeneous information
  integration. Demonstrated today via Army War
  Reserves Logistics application which integrates
  Oracle and LOGTAADS data.
• IMPACT agents can be rapidly created and deployed
  via AgentDE, the IMPACT agent development
  environment.

19
IMPACT Agent Deployment
Mobile agent execution
Agent deployment
agentRoost
AgentDE
AgentDE
AgentDE
agentRoost
State change/ message
AgentDE
AgentDE
agentRoost
20
IMPACT Support for MultiagentApplications

• AgentDE Used to build individual agents.
• IMPACT Server Provides
• yellow pages and registration services
• ontology services
• type services.
• AgentRoost Location where agent processes
  reside after the agent is built and deployed.
  Supports wake, messaging, replication, and
  mobility capabilities.
• AgentLog Logs all actions (desired by agent
  developer). Supports browsing and querying of
  actions, and rendering (some types of) actions.
  Essential for monitoring system performance and
  debugging.

21
AgentDE Agent DevelopmentEnvironment

• Built in Java. Allows rapidly building and
  testing IMPACT Agents.
• Contains libraries of actions (e.g. mobility,
  messaging) and compile-time options.
• Contains connections to different servers (e.g.
  IMPACT Servers, Oracle, Hermes, IBM Aglet).
• Built completely during Year 2 of contract.

• Proposed work in Year 3
• Extend action library.
• Extend range of servers AgentDE can connect to.
• Develop incremental action computation
  algorithms.
• Develop temporal agent programs.
• Implement agent security mechanisms within IMPACT.

22
IMPACT Agent Definition Screen
23
IMPACT AgentDE ConnectLibrary Screen
24
IMPACT AgentDE Action LibraryScreen
25
AgentRoost

• Built in Java.
• Functions supported
• houses deployed agents
• processes inter-agent and inter-roost message
  traffic
• wakes agents as appropriate
• dispatches agents to other roosts or sites if
  needed
• replicates and dispatches roosts/agents as
  needed.
• Red entries above are to be done. Green entries
  completed.

26
AgentLog

• Work done entirely in Year 2.
• Agent actions and messages are logged by the
  AgentLog.
• AgentLog allows agent developer to
• query log by content or time
• browse log
• view actions (when appropriate)
• supports playback of video, text and image
  message objects.
• Currently only supports message logging (action
  logging to be added).

27
IMPACT Server

• Preliminary IMPACT Server built in year 1 on top
  of Oracle and ThesDB.
• Year 2
• Integrated IMPACT Server with the massive PARKA
  Ontology (Hendler et al.). Provides flexible and
  scalable support for yellow pages services based
  on nearest neighbor matching.
• Conducted precision, recall and scalability tests
  on the IMPACT Server, with promising results.

28
Key scientific contribution IAgent development
theories

• Agent definition and agent programs defined in
  Year 1.
• In year 2, we developed
• Meta agent programs allow agents to reason about
  other agents states (what does the other agent
  know?) and actions (what is the other agent going
  to do?)
• Temporal agent programs allow agents to make
  commitments over time.
• Probabilistic agent programs allow agents to
  make decisions in the presence of uncertainty.
• Secure agent programs provide methods by which
  agents may provide data/services only to those
  authorized for that data/services.
• Defined formal languages and theories for all the
  above.

29
Key scientific contribution IIRegular Agent
Programs

• Year 1 We provided a detailed complexity
  analysis of agent programs - complexity is high !
• Year 2
• Identified a class of regular agents that have
  polynomial complexity (hence efficient to
  implement).
• Developed/Implemented compile time algorithm
  which associates with any agent, a set of queries
  on the agent state.
• Developed/Implemented an algorithm that computes
  what the agent is to do, given some recent state
  changes/messages.
• Developed an incremental algorithm to update what
  the agent is to do.
• Proved correctness and complexity results for all
  the above.

30
Key scientific contribution IIIAgentDE 1.0

• Designed and implemented AgentDE 1.0
• supports development of agents
• includes compile-time syntax checks
• compile-time check that associates a set of
  queries with an agent
• algorithm to compute such queries
• ran experiments testing the efficiency of agent
  evaluation
• developed (first cut) action library supporting
  agent building
• developed connections between IMPACT agents and
  other servers (e.g. Hermes, Oracle, IBM Aglet
  Tahiti server)
• AgentRoost and AgentLog built.
• Validated AgentDE on applications
• Army War Reserves logistics
• Bookstore Inventory with applications to DoD
  procurement
• IMPACT agents for network security
• JANUS Stricom simulation agents

31
Agent-based Combat InformationProcessor (AbCIP)
Vision

• CIP is an ARL product that runs tactical
  battlefield simulations.
• CIP consists of a set of servers which are
  hardwired together.
• AbCIP will
• agentize CIPs Display, Entity, DTED and
  Measures Server
• allow analysts/commanders to state conditions to
  be tracked across these servers
• automatically trigger actions by such conditions
• allow use of visualization and analytic tools.
• Ongoing Vision implemented in part,not full.

32
Agent-based Security

• ARL-Lockheed Sanders have used IBM/Aglets to
  build a network vulnerability (NV) agent
  examining .rhost files.
• IMPACT implements security across multiple
  networked machines, showing cooperation between
• IMPACT logging agents
• NV-Aglets
• This application shows how such agents can
• automatically shut off access from compromised
  machines,
• automatically notify system managers and other
  machines
• detect violations to system files.

33
IMPACT Agent Bookstore

• IMPACT Agent Bookstore shows a futuristic
  bookstore that
• has an inventory tracking agent which updates
  inventory data and triggers alerts when book
  inventories drop below specified levels
• a purchasing agent that responds to the alert and
  finds a best supplier
• an ordering agent that orders the book and
• a receiving agent that receives the books and
  updates the inventory.

34
IMPACT Janus STRICOM Simulations

• Just started.
• JANUS data involves
• dtec sensor detections file
• fires fire-event file
• kils kill-event files
• All files are binary and record events during
  simulations.
• Users can ask for email notification when certain
  conditions across these binary data sources
  occur.
• IMPACTs JANUS agents automatically provide this.

35
Research Quality Measures

• Agent book completed and to be published by MIT
  Press (expected release spring 2000).
• 31 Papers related to contract
• 14 Archival journal
• 16 Conference/workshop 1 book chapter
• Invited conference addresses
• Keynote address, Workshop on Multimedia Systems,
  Vienna, Austria, Aug. 1998.
• 5 lectures on heterogeneous info systems at
  Italian Summer School, Sep. 1998.
• Invited talk, ISMIS-99, Warsaw, Poland, July
  1999.
• Invited talk, Compulog Workshop, London, April
  1999.
• Nau etals Bridge Baron wins 1997 World Computer
  Bridge Championship
• Articles in Washington Post and New York Times

36
Army Interactions

• ARL (P. Emmerman, T. Gregory, LTC. P. Walczak)
• LIA (Bobby White, Miranda Moore)
• USMA (LTC. Marin, LTC. Byrnes, Maj. Schafer)
• STRICOM (LTC. G. Stone)
• ARL ATIRP Program Mike Gaughan (Lockheed Sanders)

37
Planned Work for Year 3

• Multiagent Development Environment (MADE)
• definition of multiagent application development
  language
• development of sound, complete and efficient
  computation algorithm
• Optimizing the status set computation algorithm
• Optimizing the incremental status set update
  algorithm
• Research on computing optimal status sets
• Research on implementing temporal agents
• Research on implementing probabilistic agents
• Research on implementing secure agents
• Implementing secure agents in AgentDE
• Implementing optimal status set computations in
  AgentDE.
• Implementing conflict resolution agents.
• Implementing visualization agents.

38
Planned Work for Year 4

• Implementing temporal agents in AgentDE
• Implementing probabilistic agents in AgentDE
• Multiagent planning applications
• MultiAgent Development Environment (MADE)
• implementation
• applications
• Complete development and deployment of AgentDE
  with applications to
• ARL CIP
• Logistics
• Electronic procurement systems
• Security