Continuous Planning and Execution

1
Continuous Planning and Execution

• Dr. Karen Myers (PI)
• Mr. David Blei
• Mr. Thomas J. Lee
• Dr. Charlie Ortiz
• Dr. David E. Wilkins
• Artificial Intelligence Center
• SRI International
• 333 Ravenswood Avenue
• Menlo Park, CA 94025
• Project URL http//www.ai.sri.com/cpef/

2
Domain Characteristics

• Tasks are complex and open-ended
• Operating environments are dynamic and possibly
  hostile
• Complete and accurate knowledge of the world can
  never be attained
• Full automation is neither possible nor desirable
• Successful operation requires a mix of
• user involvement and control
• continuous planning
• rapid response to unexpected events
• dynamic adaptation of activities

3
CPEF Foundations

• leverage several mature SRI technologies
• Procedural Reasoning System (PRS)
• Knowledge-based Reactive Control system
• SIPE-2 Hierarchical Task Network (HTN) planner
• Advisable Planner (AP)
• Multiagent Planning Architecture (MPA)
• Functional integration more than just
  interfaces ...

4
Technical Thrusts

• A. Flexible Process Management
• provide intelligent management of planning and
  execution that is responsive to the dynamics of
  the operating environment
• B. Dynamic Plan Adaptation
• provide situation monitoring, execution
  monitoring, and plan repair that enable
  reactive, timely adaptation of plans
• C. Robust Plans
• generate plans that are sensitive to the
  execution environment, knowledge limitations
• D. User Guidance
• enable users to direct and manage key aspects of
  the planning and execution processes

5
CPEF Architecture
MPA Messages
6
Infrastructure Multiagent Planning Architecture
(MPA)

• Builds on components from the Multiagent Planning
  Architecture (MPA)
• Agent-based framework for addressing large-scale
  planning problems
• distributed operations
• modularity via plug-and-play paradigm
• Protocols related to plans and planning
  activities
• layered on top of KQML
• Plan Server for storage/retrieval of plans and
  plan fragments
• Extensions to MPA for plan execution
• class of Executor agents
• protocols for plan execution, repair, updates,
  advice

7
Plan Manager Responsibilities

• Execution Tracking
• Supervise progress through execution of plans
• Knowledge Management
• Situation monitoring
• Perform information-gathering tasks
• Process Management
• Control generation of plans and options for
  outstanding tasks
• Provide timely response to user requests,
  unexpected events
• Reactive response to unexpected events
• EX downed pilot in Area of Operations
• Runtime adaptation of plans in response to
  failures, events
• EX pop-up targets, change in weather

8
Plan Manager Design

• PRS reactive control technology
• Multi-threaded, highly responsive
• Mixture of goal-directed and event-directed
  activity
• Execution tracking via Flow Model
• Await outcomes in accordance with sequencing info
  in plans
• Outcomes success, failure, unknown, time-out
• More opportunistic models would be preferable

9
PRS Reactive Control Architecture
User

• Plan partial order of goals
• Controller manages procedure application in
  accordance with Plan and New Events
• Monitors checking for critical events
• Database dynamically maintained knowledge of the
  real world
• Procedures
• goal refinement
• reactive responses

Plan
Procedures
Controller
Monitors
Database
World
10
Monitors

• Event change in the world state
• action completes, new info, time passes
• Response several possibilities
• 1. Invoke standard operating procedures
• 2. Perform plan transformations (e.g., plan
  repair, plan extension)
• 3. Record changed world model
• Monitor Classes
• Failure Monitors respond to failures that occur
  during plan execution
• Knowledge Monitors test for availability of info
  needed for decision-making
• Assumption Monitors respond to situation changes
  that violate key plan assumptions

11
Automated Monitor Generation

• Create assumption monitors with a range of
  possible responses from formal plan
  representation
• alerts, plan repairs, standard operating
  procedures
• Traversal of causal links within plan derivation
  structure to collect conditions/assumptions that
  are
• Dynamic
• Not established by earlier actions in the plans
  (ie, in initial world)
• Declared as significant
• Certain violations can be disregarded until entry
  into a critical time window (Ex weather)

PLAN Actions Monitors
12
Plan Repair

• Perform minimal-perturbation replanning for
  impacted portions of current plan
• wedge beneath failure nodes
• minimal changes provide plan continuity,
  understandability
• can be computationally expensive
• Planning-time
• Adaptation of plan in response to information
  updates
• Asynchronous Execution-time
• Adaptation of active plan during execution
• world continues to change, unaffected actions are
  executed
• Plan Manager must synchronize new plan with
  continued progress along previous plan

13
ACP Knowledge Bases

• Strategy-to-Task refinement of selected Air
  objectives
• Final plans at the level of targets, CAPS (with
  several thousand nodes)
• Key Components
• strategic knowledge for objective refinement
• hierarchical target network models
• threat models
• geographic knowledge
• force and equipment knowledge
• Assumes key intel info COGs, threats

Scope of TIE-97-1 Air Campaign Planning (ACP)
Knowledge Base
14
(No Transcript)
15
CPEF Demo Technical Highlights

• Rapid generation of alternative courses of action
• Strategy-to-task refinement of Air Superiority
  objectives
• Incremental generation of qualitatively different
  options by user (via advice)
• Application of automatically- and user-created
  monitors
• Realtime Execution Tracking in a simulated
  environment
• Asynchronous adaptation of activity in response
  to realtime monitoring of
• situation changes (Ex Downed Pilot)
• plan execution results (Ex failure to
  neutralize critical targets)
• Advised Plan Repair

16
Conclusions

• CPEF Prototype demonstrates flexible process
  management for living plans in Air Campaigns
• Full Spectrum plan generation, execution, repair
• Several Notable Technical Accomplishments
• Models for Plan Management, Execution Tracking
• Generalized Failure Models, Repair Methods
• Promising preliminary work on Open-ended Planning
• Major Contributor within the JFACC Program

CPEF
17
Backup Slides
18
Execution Models

• Direct Execution (Do it!)
• actions are dispatched directly by the system
• EX controller for a mobile robot
• Indirect Execution (Supervisory)
• plan is executed by diverse, distributed agents
• agents are pre-assigned execution tasks
• status of action execution is not directly
  available
• delays in redirecting agents that perform planned
  actions
• time lag on receipt of information about the world

19
Generalized Failure Models

• Limited scope of current models
• Precondition Failure action precondition not
  satisfied
• Action Failure intended effects not achieve
• Maintenance Failure established condition no
  longer maintained
• New directions --- beyond plan dependency
  structures
• Unattributable Failure no individual action has
  failed or assumption violated yet plan is deemed
  inadequate
• Ex CA indicates failure to establish
  sufficiently strong breach of IADS
• Aggregate Failure require collections of
  failures, possibly with key relationships among
  them (eg, A fails then B fails)
• Ex key subset of a target network

20
PRS Control Loop
Execution Cycle 1. New information arrives that
updates facts and goals 2. Acts are triggered by
new facts or goals 3. A triggered Act is
intended 4. An intended Act is selected 5. That
intention is activated 6. An action is
performed 7. New facts or goals are posted 8.
Intentions are updated
21
CPEF Continuous Planning and Execution
Framework

• process management technology for living plans
• plan creation, execution, repair
• vertical slice of the JFACC system

JFACC System
Layered View of CPEF
Workflow Management
C P E F
Plan Gen
Specialized Components
22
Process Management Generality and Ubiquity
Process Management
Process Management
Process Management
Info Needs
Agents
Plans
Air Campaigns
ISR
CPEF (JFACC, SUO)
SWIM (AIM)
TRAC (CoABS)
23
Accomplishments Technical

• Process Management for plan generation, indirect
  execution, monitoring, repair
• Automated extraction of monitors from plans
• Generalized models of failure and execution
  monitoring
• Mixed-initiative options generation and plan
  repair (using advice)
• Preliminary models for open-ended planning
• Towards a Framework for Continuous Planning and
  Execution, AAAI 1998 Fall Symposium on
  Distributed, Continual Planning (Special Issue of
  AI Magazine)