IMPACT: Future Directions
(years 3 and 4)
V.S. Subrahmanian
University of Maryland
1
Research Directions

“Smart” Agent Development:





MultiAgent Development Environment (MADE)




Implementing Temporal and Uncertain Agent Reasoning
Agent Security Toolkit
Spatial Agent Programs
Specialized agent activities: planning, learning .
How can multiple agents developed in AgentDE work together?
How can such “working together” be configured within IMPACT?
Study interaction protocols and properties of multiagent applics.
Applications:



Full agentization of CIP Alert Application
Full-fledged logistics application
Full fledged E-store application
1999 IMPACT Workshop
2
Implementing Temporal and
Uncertain Agent Reasoning
DONE TO DATE
 In year 2, we proposed
temporal agent programs
(TAPs), and probabilistic agent
programs (PAPs).
 Developed a semantics for
them and developed
algorithms to compute TAPs.
TO BE DONE IN YEARS 3-4
 Extend AgentDE to handle
temporal agent programs.
 Implement TAP status set
computation algorithms.
 Develop probabilistic status set
computation algorithms and
prove correctness theorems.
 Extend AgentDE to handle
probabilistic agent programs.
 Implement PAP status set
computation algorithms.
1999 IMPACT Workshop
3
Agent Security Toolkit

In year 2, we proposed a theoretical framework for
agents to




articulate their security requirements and
“protect” themselves from security violations.
AgentDE will include an Agent Security Toolkit (AST).
AST will contain the following capabilities:





agent approximation methods (how does agent A approximate
agent B and update its approximation?)
agent security specification programs (what does agent A want to
protect from an arbitrary agent B?)
history logging
coupling agent programs and security structures.
security preservation algorithms
1999 IMPACT Workshop
4
Spatial Agent Programs


Agents must have the ability to reason about space-time.
We plan to develop the notion of a spatio-temporal
agent program (STAP).





Syntax of STAPs
Formal semantics of STAPs, extending TAPs
Coupling STAP implementations to scalable spatial data
structures.
Developing STAP-feasible status set computation algorithms and
proving their correctness.
Implementing STAPs as an extension of the AgentDE.
1999 IMPACT Workshop
5
Planning agents


Many (but not all) agents need
to plan.
Planning capabilities can be
incorporated within agents in
many ways:


Planning can be viewed as an
IMPACT action.
IMPACT action rule heads can
be extended to include:
– hierarchical task networks
in rule heads instead of
single actions.
– Sets of actions with
attached temporal
constraints.

We plan to:




develop a semantics for
agent programs with these two
modes of planning
study the complexity and
expressive power of these
two alternatives
develop implementation
methods and conducted
experiments
Build applications of multiagent
planning in the areas of:


noncombatant evacuation
ops
logistics planning
1999 IMPACT Workshop
6
MultiAgent Development
Environment (MADE)



AgentDE allows agent
developers to build agents that
interact with other agents.
However, when a set of
interacting agents are defined,
one by one, it is difficult to
evaluate how this multiagent
application will behave over
time.
MultiAgentDE will contain the
theory and algorithms needed
by agent developers to ensure
that their multiagent application
preserves a variety of
important application specific
conditions.
WORK TO BE DONE
 Develop a language to query
a multiagent application built in
IMPACT.




Querying over alternative
“future” worlds
Evaluation of integrity
constraints over all possible
futures.
Develop techniques to
implement this language.
Implement this language, and
define tools to allow agent
developers to interact with their
application.
1999 IMPACT Workshop
7
IMPACT Agent Library








Consists of a set of COTS-tools based agents that
different (more sophisticated) civilian and military
applications may want to extend.
“Database” Agents: Oracle, Access, ObjectStore.
Spreadsheet Agents: Excel.
“Media” Agents: Image, Video, Free text.
Language Agents: Machine translation agents.
Planning Agents: SHOP Hierarchical Task Network
Agents.
GIS and Map Management Agents.
Web Topic Agents: Weather, Map, etc.
1999 IMPACT Workshop
8
Applications: Agent-based Combat
Information Processor (AbCIP)



CIP is currently a set of servers.
We propose to make each of these servers into an
IMPACT Agent.
We propose to develop new agents providing a wide
variety of “active” CIP-monitoring and analysis services.
Users register conditions of interest.
 Users describe actions to be taken when the conditions of
interest satisfy some conditions.
The CIP Multiagent application will include a mix of CIP-agents, GIS
agents, database agents, and planning agents.


Preliminary work Started. Demo in Room 2120
1999 IMPACT Workshop
9
Applications: Multi Agent
Logistics System (MALS)



Current IMPACT agents have the ability to access large
scale logistics database and trigger actions.
MALS will support the “track inventory status => create
plan for resupply => execute plan => track plan” cycle.
MALS will include a mix of:





Current logistics inventory agents
Excel agents
Financial tracking and ordering agents
Planning agents
GIS Agents
1999 IMPACT Workshop
1
Conclusions





AgentDE 1.0 is near-ready for
the deployment of single and
multiagent applications.
We expect to provide “beta”
versions of AgentDE 1.0
(without documentation) for
testing to ARL in Sep. 1999
and with documentation in
Nov. 1999.
AgentDE 2.0 =>
Temporal/Spatial
AgentDE 3.0 => Probabilistic,
planning.
AgentDE 4.0 => Security

Multiagent Development
Environment (MADE)




Query Language
Algorithms
Implementation
Applications



AbCIP
MALS
Electronic store
1999 IMPACT Workshop
1