MPA for the SUO Planning and Decision Aid

1
MPA for the SUO Planning and Decision Aid
David E. Wilkins SRI International
Artificial Intelligence Center email
wilkins_at_ai.sri.com URL http//www.ai.sri.com/wi
lkins June 1999
2
Long-term Contributions of PDA to the Soldier

• Shared plan representations cognizant of miltary
  principles, doctrine, SOPs and TTPs. Makes
  possible
• Doctrinally correct plan usable by everyone.
• Fast generation of multiple distinct COAs,
  including ones the commander may not have
  considered.
• Support commander creativity through human
  guidance.
• Uniformly high plan quality, even during
  high-stress crises.
• Continuous analysis of the plan using multiple
  metrics.
• Monitor plan execution and respond quickly to
  events, helping the commander modify the plan
  appropriately.
• Provide relevant information to other echelons,
  allowing fast communication while preserving
  bandwidth

3
Multiagent Planning Architecture (MPA)

• Agent-based framework (efficiency and modularity)
• Enable interoperation of diverse planning
  technologies
• plan generation, scheduling, temporal reasoning,
  evaluation, ...
• Provide a shared plan representation
• Provide organizational structure, control
  strategies, and communication protocols,
  wrappers
• Develop planning and metaplanning agents
  specifically for collaborative planning and
  scheduling
• Support dynamic, event-driven reconfiguration of
  planning organization and strategies at runtime

4
MPA Components

• Planning Cells - organizational units for agents
• Planning Agent (PA), Meta Planning Agent
  (Meta-PA)
• Planning-Cell Managers
• Shared plan model and representation
• based on extenstions to Act formalism
• Communication Protocols
• Transport Level robust, reuse existing
  technology
• Content Level high-level performatives
• Plan Server - central repository for plans and
  related information
• Process Management - extensible set of reactive
  control policies for implementing problem-solving
  strategies
• Support- documentation of each agents I/O,
• Wrappers for individual technologies

5
Communication

• Messages use KQML-like performatives
• communication performatives
• plan performatives
• MPA wrappers and libraries for
• message passing
• multithreaded processing
• tracing and logging of messages
• Communication substrates
• KQML (Lockheed-Martin and UMBC)
• ILU (Xerox PARC)
• OAA (SRI International)

6
Examples of MPA messages

• Requesting solutions from the meta-PCM
• (evaluate content (multiple-solve
• task air-superiority advice-contexts
• ((yuma (ingress-at-yuma
  deny-air-picture))
• (yuma-sf (ingress-at-yuma
  ingress-2-at-san-francisco
• 
  breach-at-two-places)))))
• Retrieving the plan from the plan server
• (ask-all content (query-plan
• task air-superiority plan
  plan-yuma-sf view ascii)

7
MPA Plan Server

• Central repository for planning information
• Hierarchical model of plans (plans, tasks, and
  action networks)
• Answers queries about the plan, providing
  multiple views
• Supports a broad range of plan management
  capabilities
• Annotations - declarations of high-level
  attributes of plans, planning process
• Product Annotation pedigree, flaws, plan
  quality, resource status
• Process Annotation time spent on plan
  components, current status
• Triggers - rules used to notify cell agents of
  planning events
• Example trigger notifies PCM of overutilized
  resource annotation,
• PCM reacts by changing planning strategy

8
MPA Applied to ACP

• Validate MPA by integrating several systems in
  DARPA Planning Initiative (TIE 97-1)
• INSPECT (ISI) OPIS (CMU)
  Advisable Planner (SRI)
• SIPE-2 (SRI) ACS (UMass)
  Process Panel (AIAI -UEdin)
• APAT (ISX) VISAGE (MAYA)
• Domain is Air Campaign Planning
• thousands of objects, several thousand nodes in
  each plan
• plan down to support mission level (must allocate
  supporting resources)
• air superiority objective only
• targets grouped into networks which depend on
  other networks
• network effectiveness is modeled quantitatively

9
TIE 97-1 Architecture
MPA Planning Agents
MPA Agent Evaluation Toolkit
Process Monitor
APAT GUI Plan Viewer
ACS
VISAGE
Advisable Planner (SIPE-2) Planner
(SIPE-2) Scheduler (OPIS) Planning-Cell Manager
(PRS)
Inspect
10
TIE 97-1Demonstrations

• Sept 98 - EFX 98, Ft. Walton Beach FL
• May 98 - ARPI Workshop, Monterey CA
• Feb 98 - DARPA, Arlington VA
• Dec 97 - JFACC PMR, San Pedro CA
• Nov 97 - ARPI Workshop, San Francisco CA

Increasing Capabilities
11
Inside an MPA Planning Cell
Create a plan
PA GUI/Advice Manager (AP and PRS)
Plan Complete
Meta-PA Planning-Cell Manager (PRS) Planning-Cell
Designator
Create a plan
Request Plan ok?
Plan Complete
or
PA (meta-PA) Critic Manager all (SIPE-2)
Request Expand Next Level
PA Temporal Reasoner (Tachyon)
Cue
ACT2
(TEST (ready unit1))
Cue
ACT1
Resource Critic
PA Scheduler (OPIS)
PA Temporal Reasoner (OPIS)
PA Schedule Critic (new)
PA (meta-PA) Search Manager one-level (SIPE-2)
OPIS
PA Temporal Constraint Critic (SIPE-2)
KQML
Agent
Message
SIPE-2
SRI International
12
MPA PDA Cell
Higher Echelon Plan Complete
Create a plan
Company Commander
Cell Manager (CPEF) Cell Designator
New Op Order
Request Modify plan
Plan Complete
Planner (SIPE-2, O-Plan)
Cue
ACT2
(TEST (ready unit1))
Cue
ACT1
Resource Critic
At WTA
Terrain Reasoner
Fire Control Planner (e.g.)
KQML
Agent
Message
SRI International
13
Multiple PDA Cell Configuration
Higher Echelon Cell Manager (CPEF)
Plan
Plan
PDA Cell
PDA Cell
New Order
New Order
Possibly Shared Agents
KQML
SRI International
Message
14
Opportunities Enabled by MPA

• Facilitate collaboration
• agents exchange information and influence each
  other during planning
• can easily explore different degrees of
  collaboration
• Can more easily explore/evaluate
• different organizational units for flexible
  control policies
• different planning styles and strategies
• new or alternative technologies

15
Backup Slides
16
The Problem
Support rapid, accurate military decision making
in information-rich warfighting environment
17
Example MPA Performatives
annotation Insert Delete
Ask-All Ask-One trigger Insert
Delete
Ask-All Ask-One update-task Tell
Delete update-plan Tell
Delete query-task
Ask-all Ask-One query-plan Ask-All
Ask-One query-node Ask-All Ask-One ping
Evaluate pcd Tell
Performatives accepted by the Plan Server
18
Planning Cells

• Hierarchically organized collection of planning
  agents
• Composed dynamially by the Planning-Cell Manager
• Planning Cell Designator (PCD) provides registry
  of agents to fill various roles
• Planner SIPE-2 Scheduler OPIS
• Distribute tasks and PCD to PAs and meta-PAs
• Each cell includes a plan server

Planning cell
Planning-Cell Manager
19
Planning Cell Manager (PCM)

• Persistent agent continuously accepting tasks
• Oversees problem-solving within a Planning Cell
• Provides a range of problem-solving behaviors and
  strategies
• user-configurable
• can adapts strategies in response to runtime
  activities
• PRS-based PCM
• PCD and process state encoded in PRS database
• strategies encoded in Acts
• Example strategy expand and critique by levels
• implemented by 14 Acts
• small library of different PCMs
• various cell configurations, problem-solving
  strategies

20
Meta Planning-Cells

• Planning cell composed of a set of planning cells
• A Meta-PCM controls and coordinates the cells
• Accepts multiple planning requests, including
  advice
• Distributes tasks to available planning cells
• Gathers results for possible comparison

Meta Planning-Cell
Meta Planning-Cell Manager
Planning Cell
Planning Cell
Planning Cell
21
Act Plan Server

• PRS-based Plan Server in current MPA applications
• builds on reactive control capabilities
• grounded in the Act Formalism model for plans
• Task problem to be solved, advice, assumptions
• Action Networks partial orders of activities, at
  multiple abstraction levels
• Plan linked sets of action networks
• Nodes individual activites (goals, actions)
• limited persistence, versioning, browsing, and no
  access control

22
MPA Multiple Planning-Cell Configuration
Meta Planning- Cell Manager (PRS)
task solved
task solved
Planning Cell
Planning Cell
request solution to task
request solution to task
Shared Agents
Shared Agents
KQML
SRI International
Message