+ All Categories
Home > Documents > Mae 345 Lecture 23

Mae 345 Lecture 23

Date post: 10-Apr-2018
Category:
Upload: algefer
View: 217 times
Download: 0 times
Share this document with a friend

of 12

Transcript
  • 8/8/2019 Mae 345 Lecture 23

    1/12

    Task Planning andMulti-Agent Systems

    Robert StengelRobotics and Intelligent Systems, MAE 345,

    Princeton University, 2009

    Copyright 2009 by Robert Stengel. All rights reserved. For educational use only.

    http://www.princeton.edu/~stengel/MAE345.html

    Decision making

    Task decomposition, communities, and connectivity

    Cooperation, collaboration, competition, and conflict

    Multi-agent architectures

    Control Functions ofan Intelligent Agent

    Conscious Thought

    - Awareness- Focus

    - Reflection

    - Rehearsal- DeclarativeProcessing of Knowledge or Beliefs

    Unconscious Thought

    - Subconscious Thought

    > ProceduralProcessing> Communication

    > Learned Skills

    > Subliminal Knowledge Acquisition- Preconscious Thought

    > Pre-attentive DeclarativeProcessing

    > Subject Selection for Conscious Thought> Concept Development

    > Information Pathway to Memory

    > Intuition

    Intelligent AgentCharacterized byDeclarative, Procedural,and Reflexive Actions

    ReflexiveBehavior- Instantaneous Response to Stimuli

    - Elementary, Forceful Actions

    - Stabilizing Influence

    - Simple Goals

    Task Planning Goals Accomplish an objective

    Make a decision

    Gather information

    Build something

    Analyze something

    Destroy something

    Determine and follow a path

    Minimize time or cost

    Take the shortest path

    Avoid obstacles or hazards

    Work toward a common goal

    Integrate behavior with higherobjectives

    Do not impede other agents

    Central Pacific and Union Pacific Railroads

    meet in Promontory, Utah, 1869

  • 8/8/2019 Mae 345 Lecture 23

    2/12

    More Task Planning Goals

    Provide leadership forother agents

    Issue commands Receive and decode

    information

    Provide assistance toother agents

    Coordinate actions

    Respond to requests

    Defeat opposing agents Compete and win

    Common Threads inTask Accomplishment

    Optimize a cost function

    Satisfy or avoid constraints

    Exhibit desirable behavior Tradeoff individual and team goals

    Use resources effectively and efficiently

    Negotiate

    Cooperate with team members

    Overcome adversity and ambiguity

    TaskPlanning

    Situation awareness

    Decomposition and identification of communities

    Development of strategy and tactics

    PhaseProcess Outcome

    Objective Tactical(short-term)

    SituationAssessment

    SituationAwareness

    Strategic(long-term)

    Comprehension Understanding

    Boyd!s OODA Loop

    for Combat Operations

    Derived from air-combatmaneuvering strategy

    General application to learning

    processes other than military

  • 8/8/2019 Mae 345 Lecture 23

    3/12

    Endsley, 1995

    Elements ofSituation Awareness

    Perception

    Comprehension

    Projection

    Important Dichotomiesin Planning

    Strength, Weakness, Opportunity, and

    Threat (SWOT) Analysis

    Knok-Knoks and Unk-Unks

    Simultaneous Locationand Mapping (SLAM) Build or update a local map

    within an unknown environment Stochastic map, defined by mean

    and covariance

    SLAM Algorithm = State estimationwith extended Kalman filter

    Landmark and terrain tracking

    Durrant- Whyte et al

    Multi-Agent Control Example Based on

    Linear-Quadratic-Gaussian

    (LQG) Optimal Control

    E(J) = E

    ! x(tf )"# $% + L x(t),u(t)[ ]to

    tf

    & dt

    =1

    2x

    T(tf )S fx(tf )+ x

    T(t)Qx(t) + u

    T(t)Ru(t)"# $%

    to

    tf

    & dt'()

    *)

    +,)

    -)

    '

    (

    ))

    *

    )))

    +

    ,

    ))

    -

    )))

    Quadratic cost function

    Linear dynamic model

    !x(t) = Fx(t) +Gu(t) + Lw(t)

  • 8/8/2019 Mae 345 Lecture 23

    4/12

    Stochastic Optimal Control Law

    u(t) = !R!1G

    TS(t)x(t) = !C(t) x(t)

    S(t) = !FTS(t)!S(t)F + S(t)GR

    !1G

    TS(t)!Q, S(tf) = Sf

    Solution of Euler-Lagrange equationsleads to optimal feedback control law

    Where S(t)is the solution to a matrixRiccati equation

    (Estimator not shown)

    A Federated Optimization Problem

    x(t) = Fx(t)+Gu(t) =FA

    FB

    A

    FAB

    FB

    "#

    %&xA

    xB

    !

    "#

    $

    %&+

    GA

    GB

    A

    GAB

    GB

    "#

    %&uA

    uB

    !

    "#

    $

    %&

    Dynamic models for two agents, A and B, are coupled to

    each other and expressed as a single system

    E(J) = E1

    2xT(t)Qx(t)+ u

    T(t)Ru(t)!" #$

    to

    tf

    % dt&'(

    )(

    *+(

    ,(

    = E1

    2xAT

    xBT!

    "#$

    QA QBA

    QAB

    QB

    !

    "

    --

    #

    $

    .

    .

    xA

    xB

    !

    "--

    #

    $..+ uA

    TuB

    T!"

    #$

    RA RBA

    RAB

    RB

    !

    "

    --

    #

    $

    .

    .

    uA

    uB

    !

    "--

    #

    $..

    !

    "

    --

    #

    $

    .

    .to

    tf

    % dt&

    '(

    )(

    *

    +(

    ,(

    u(t) = !Cx(t) =uA

    uB

    "

    #

    $$

    %

    &

    ''= !

    CA

    CB

    A

    CA

    BC

    B

    "

    #

    $$

    %

    &

    ''

    xA

    xB

    "

    #

    $$

    %

    &

    ''

    Cost function minimizes performance-control tradeoff

    Optimal feedback control laws are coupled to each other

    A Distributed Optimization Problem

    x(t)= Fx

    (t)+Gu

    (t)=

    FA

    0

    0 FB

    "# %&

    xA

    xB

    "# %&+

    GA

    0

    0 GB

    "# %&

    uA

    uB

    "# %&

    Each sub-system can be optimized separately

    Each control depends only on separate sub-state

    E(J) = E1

    2xT(t)Qx(t) + uT(t)Ru(t)!" #$

    to

    tf

    % dt&'(

    )(

    *+(

    ,(

    = E1

    2xAT

    xBT!

    "#$

    QA

    0

    0 QB

    !

    "--

    #

    $..

    xA

    xB

    !

    "--

    #

    $..+ uA

    T uBT!

    "#$

    RA

    0

    0 RB

    !

    "--

    #

    $..

    uA

    uB

    !

    "--

    #

    $..

    !

    "

    --

    #

    $

    .

    .to

    tf

    % dt&'(

    )(

    *+(

    ,(

    u(t) = !R

    A0

    0 RB

    "

    #

    $$

    %

    &

    ''

    !1

    GTSx(t) = !Cx(t) =

    uA

    uB

    "

    #

    $$

    %

    &

    ''= !

    CA

    0

    0 CB

    "

    #

    $$

    %

    &

    ''

    xA

    xB

    "

    #

    $$

    %

    &

    ''

    Coupling between actions of two agents, A and B, is negligible

    Pursuit-Evasion:A Competitive

    Optimization Problem

    Linear model with two competitors, Pand E

    x(t) = Fx(t)+Gu(t) =x

    P

    xE

    !

    "#

    $

    %&=

    FP

    0

    0 FE

    !

    "#

    $

    %&x

    P

    xE

    !

    "#

    $

    %&+

    GP

    0

    0 GE

    !

    "#

    $

    %&uP

    uE

    !

    "#

    $

    %&

    Pursuer!s goal: minimize final miss distance

    Evader!s goal: maximize final miss distance

    Example of a differential game, Isaacs (1965), Bryson & Ho (1969)

  • 8/8/2019 Mae 345 Lecture 23

    5/12

    Pursuit-Evasion:A Competitive

    Optimization Problem

    Quadratic minimax (saddle-point) cost function

    Optimal control laws for pursuer and evader

    E(J) = E1

    2xT(t

    f)S(t

    f)x(t

    f)!" #$ +

    1

    2xT(t)Qx(t) + u

    T(t)Ru(t)!" #$

    to

    tf

    % dt&'()(

    *+(

    ,(

    = E

    1

    2xPT(t

    f) xE

    T(tf)!

    "-#$.

    SP SPE

    SEP SE

    !

    "

    --

    #

    $

    .

    .f

    xP (tf )

    xE(tf)

    !

    "

    --

    #

    $

    .

    .

    +1

    2x

    P

    T(t) x

    E

    T(t)!

    "#$

    QP QPE

    QEP QE

    !

    "--

    #

    $..

    xP (t)

    xE(t)

    !

    "--

    #

    $..+ u

    P

    T(t) u

    E

    T(t)!

    "#$

    RP 0

    0 /RE

    !

    "--

    #

    $..

    uP(t)

    uE(t)

    !

    "--

    #

    $..

    !

    "

    --

    #

    $

    .

    .to

    tf

    % dt

    &

    '

    (((

    )

    (((

    *

    +

    (((

    ,

    (((

    u(t) =u

    P(t)

    uE(t)

    !

    "

    ##

    $

    %

    &&= '

    CP(t) C

    PE(t)

    CEP

    (t) CE(t)

    !

    "

    ##

    $

    %

    &&

    xP(t)

    xE(t)

    !

    "

    ##

    $

    %

    &&

    Strategy/Tactics Developmentand Deployment

    Development of long- and short-termactions/activities for implementation and

    operation Sequence of procedures to be executed

    fixed or adaptive

    Exposition of approach Rules of engagement

    Concept of Operations (CONOPS)

    Spectrum of flexibility Rigid sequence Learning systems

    Think Expert System

    Program Management:Gantt Chart

    Project schedule

    Task breakdown and dependency Start, interim, and finish elements

    Time elapsed, time to go

    Program Evaluation and ReviewTechnique (PERT) Chart

    Milestones

    Path descriptors

    Activities, precursors, and successors

    Timing and coordination

    Identification of critical path

    Optimization and constraint

  • 8/8/2019 Mae 345 Lecture 23

    6/12

    Path Planning

    Trajectory decompositionand segmentation

    Environment idealizationand nominal path

    Waypoints

    Path primitives (line, circle, etc.)

    Timing and coordination

    Obstacle detection and avoidance

    Feasibility and regulation

    Optimization and constraint

    Task Decomposition: Path Planning

    Tessellation (tiling) of a2-D decision space

    Given set of points, e.g.,obstacles, destinations,or centroids of multiplepoints

    2-D Voronoi diagram Polygons with sides

    equidistant to two nearestpoints (black dots)

    Distance= Euclidean norm other metrics can be used

    transformations

    Voronoi diagram

    Dual Graph is theDelaunay Triangulation

    Edges (black) connect all triplets of points lying on

    circumferences of empty circles, i.e., containing noother points

    Voronoi segment boundaries (red) are perpendicularto each edge

    Minimum spanning treeis a subset of theDelaunay graph

    Paths with farthestdistances from obstacles

    Voronoi Diagrams inPath Planning

    Obstacle avoidance

    Nearest neighbor identification, e.g., closesthospital

  • 8/8/2019 Mae 345 Lecture 23

    7/12

    Voronoi Diagrams inData Processing

    Computer graphics textures (2-D and 3-D meshes)

    Voronoi diagram with

    fuzzy boundaries

    http://www.data-compression.com/vqanim.shtml

    Vector quantization in data compression

    Density characterization (3-D mesh)

    2004 DARPAGrand

    ChallengeRoute

    Specification

    Waypoint # Latitude Longitude

    Lateral

    Boundary

    Offset

    Speed Limit

    (mph)

    Drop-Dead

    Time,

    hours minutes seconds1 34.7607233 -117.0107599 90 10 #### #### ####2 34.7607515 -117.0107413 15 10 #### #### ####3 34.761033 -117.01056 12 10 #### #### ####4 34.7611515 -117.0105663 12 10 #### #### ####5 34.7612679 -117.0106085 10 10 #### #### ####

    732 34.9313317 -116.7157473 13 15 13 30 0

    2582 35.6168352 -115.3820928 10 5 #### #### ####2583 35.6164206 - 115.382102 10 5 #### #### ####2584 35.6162642 -115.3821041 10 5 #### #### ####2585 35.6162367 -115.3821032 6 2 #### #### ####2586 35.6162012 -115.3821023 12 0 #### #### ####

    Task Decomposition:Community Identification

    Connectivity ofindividuals

    Individualsassemble incommunities orclusters

    Complexnetworks

    Randomnetworks

    Small-worldnetworks

    Scale-freenetworks

    Degrees ofseparation

    Fully connectedRandom

    Clustered small worldSmall world ring lattice

    Community Communication

    Strogatz, 2001

    Scale-Free Networks Frequency and cumulative distributions of cluster sizes inversely

    proportional to !x

    Strogatz, 2001

    Scale-Free

  • 8/8/2019 Mae 345 Lecture 23

    8/12

    Community Examples Associations

    Governments Agencies

    Laboratories Managers

    Scientists

    Military organizations Army

    Corps Division

    Brigade

    Regiment Battalion

    Company

    Platoon

    Squad Soldier

    Special Operations

    Terrorist organizations

    Families

    Classmates

    Neighbors

    Social Networks

    Facebook LinkedIn

    Media Networks

    Corporations

    Employees

    Customers

    Sports Leagues Teams

    Managers

    Players

    Trainers

    Airlines

    Cities

    Multi-Agent Systems

    Specialized vs. general-purpose agents

    Organizational models

    Cooperators

    Leader/follower (hierarchical)

    Equal members Collaborators

    Air, ground, and sea traffic

    Customers

    Competitors

    Individual game players

    Sports teams

    Political/military organizations

    Negotiators

    Politicians

    Employer/employee representatives

    Multi-Agent Systems

    Cooperation and collaboration should

    lead to win-win (non-zero-sum)solutions

    Competition should lead to win-lose(zero-sum) solutions

    Negotiation should lead to win-win butmay lead to win-lose solutions

    Team Work

    Cooperativemaneuvering

    Collaborativemaneuvering

  • 8/8/2019 Mae 345 Lecture 23

    9/12

    Typical Characteristics of Multi-Agent Architectures

    Federated (centralized)problem solving

    Doctrinaire Coupled

    Synchronous

    Fragile

    Complex

    Strategic

    Information-rich

    Unified

    Integrated

    Top-down

    Distributed problemsolving

    Autonomous Independent

    Asynchronous

    Robust

    Simple

    Tactical

    Parsimonious

    Idiosyncratic

    Modular

    Bottom-up

    Hierarchical Tree orHub-and-Spoke Network?

    What is the Nature, Quality, andSignificance of Connections?

    CommunicationCollaboration

    Coordination

    Negotiation

    Competition

    Conflict

    Connections May ConnoteDifferent Relationships

    Communication Collaboration

    Coordination

    Negotiation

    Competition

    Conflict

  • 8/8/2019 Mae 345 Lecture 23

    10/12

    A CooperativeMulti-Agent System

    Air Traffic Management: ACollaborative Multi-Agent System

    http://www.natca.org/flight-explorer/united-states.aspx

    Collaboration Coordination

  • 8/8/2019 Mae 345 Lecture 23

    11/12

    Competition Principled Negotiation:Getting to Yes(Fisher, Ury, 1981)

    Separate the agents from the problemFocus on interests, not positions

    Invent options for mutual gain

    Insist on using objective criteria

    Principled Negotiation:Getting Past No(Ury, 1991)

    Prepare by identifying barriers to cooperation, options, standards,and your Best Alternative to a Negotiated Agreement (BATNA)

    Understand your goals, limits, and acceptable outcomes Buy time to think

    Know your hot buttons, deflect attacks

    Acknowledge opposing arguments

    Agree when you can without conceding

    Express your views without provoking

    I statements, not you statements

    Negotiate the rules of the game

    Reframe the negotiation

    Build a golden bridge that allows opponent to retreat gracefully

    Engage third-party mediation or arbitration

    Aim for mutual satisfaction, not victory

    Forge a lasting agreement

    Conventional Conflict

  • 8/8/2019 Mae 345 Lecture 23

    12/12

    Terrorist Attack

    http://firstmonday.org/htbin/cgiwrap/bin/ojs/index.php/fm/article/view/941/863

    The 9/11 TerroristNetwork

    Control Functions ofan Intelligent Agent

    ConclusionMAE 345 Course Learning Objectives

    ! Dynamics and control of robotic devices.! Cognitive and biological paradigms for system design.! Estimate the behavior of dynamic systems.! Apply of decision-making concepts, including neural networks, expertsystems, and genetic algorithms.! Components of systems for decision-making and control, such assensors, actuators, and computers.! Systems-engineering approach to the analysis, design, and testing ofrobotic devices.! Computational problem-solving, through thorough knowledge,application, and development of analytical software.! Historical context within which robotics and intelligent systems haveevolved.! Global and ethical impact of robotics and intelligent systems in thecontext of contemporary society.! Oral and written presentation.


Recommended