Intelligent Systems Intelligent systems AI Programming course TDDD10 AI Programming

advertisement
IntelligentSystems
TDDD10AIProgramming
Introduction
CyrilleBerger
Warehousetransport Robotictruckconvoy Automaticparking
2/51
Dronedeliverysystem Longdistancedelivery Robotvacuumcleaner
Intelligentsystems
Soccerteams(RoboCup)
Fastrobots(SICKChallengex)
AIProgrammingcourse
RescueRobots(RoboCup)
MultiAgentRescue
Simulation(RoboCup)
All-terrainnavigation(TechXchallenge)
3/51
AIProgramming
Coursegoal
Programmingofintelligent
autonomoussystems
ImplementationandintegrationofAI
techniquesintoworkingsystems
Focusonreal-worldenvironments
usinganagent-oriented
paradigm
Gainexperiencewithsomeofthe
difficultieswhenbuildingreallargescaleAIsystems
Giveanintroductiontoagent-orientedAI
programmingthroughtheuseofasearch
andrescuesimulator(RoboCup).
Introducesfundamentalproblemsand
techniquesrelatedtoconstructingagentorientedAI-systems.
Gainpracticalexperienceinsolvingthese
problemsandimplementing,integrating,
andevaluatingthesetechniques.
5/51
6/51
Courseorganization
Divisionoftime
Listoflectures
AIProgramming:
Introductionto
3
AgentsandAgents
4
Communication
5
MultiagentDecision
6
CooperationAndCoordination
7
CooperationAndCoordination
8
Machine
9
Knowledge
10
PuttingItAll
1
2
20hlectures(in10sessions)
60hlabs(in15sessions)
12hseminars(in4sessions)
68hhomework
Workrepartition
Labsinagroupofthreeforamonth
Implementationofamultiagentsystemsina
groupof8-10
StudyandimplementationofanAItechnique
7/51
8/51
RoboRescueSimulator
Listoflabs
PracticalIntroductiontoRoboCup
RescueSimulation
2
ImplementingefficientPathPlanning
3
CommunicationandContractNet
Protocol
4
Exploration
5
Taskallocation
6
MachineLearning
1
9/51
10/51
Labs
Deadlinesanddeliverables
Registeryourgroup(3studentsper
group)inWebReg:https://
www.ida.liu.se/webreg/
Followlabinstructionsat:
https://www.ida.liu.se/~TDDD10/
info/labs.en.shtml
Onceyouaredonewithalab:
October,10th:labsarefinished
October,29th:individualassignmentprojectis
selected
November,19thor/and26th(preliminary):
presentationofindividualprojectduringa
seminar
December,15th(preliminary):presentationof
theteam
December,15th(preliminary):draftofreports
aresubmited
January,6th:finalteamandindividualreport
Commitin
Presentyourresultstothelab
11/51
12/51
Courseexamination
Requirements
Labs1to4
Individualreport
Groupreport
Grade
grade=0.75*individual_grade
+0.25*group_grade
+0.5*lab_5_done
+0.5*lab_6_done
13/51
CourseMaterial
Courseimprovements
Thesimulatorisnotrunningverywell.
Goodnews:wegotnewlab
Badnews:someoftheroomsarestillequipedwiththeold
Sinceitisjava,theyshouldrunonWindowsandOSX,
instructionstocomelater
TheAPIisbigandconfusing
newlecture(on
Ittakestoomuchtimetogetresults
livewithit,intherealworlditcantakedaysbeforeyougeta
butinthenewlecture,wewillseehowtocreatesimple
Ihavespenttoomuchtimeontheindividualassignmentanddid
nothavetimeforthegroupassignment,andvice-versa
Individualassignmentispartofthegroupassignment.
14/51
TeachingPhilosophy
Iamnotperfect,Idon'tknoweverything.Iam,
alas,human.
Icannotteachyouanythingbyforce.However,I
cangiveyouaframeworkforyourownlearning
experiencebysettingboundariesanddefining
requirements.
Withoutyourfeedbacknoimprovement.
Learningisafunadventure!
Thecourseisgoingtobewhatyouwant
tomakeofit.Participate,beactiveand
involved.
AnIntroductionto
MultiAgentSystemsby
MichaelWooldridge,
ISBN:9780470519462
MultiagentSystemsby
GerhardWeiss,ISBN
978-0-262-01889-0
andthelectures
slides(eventually)
15/51
16/51
WhatareMulti-AgentSystems?
AnMAScanbedefinedasalooselycouplednetworkof
problemsolversthatinteracttosolveproblemsthatare
beyondtheindividualcapabilitiesorknowledgeofeach
problemsolver(DurfeeandLesser1989)
Multi-AgentSystems
Theseproblemsolvers,often
calledagents,areautonomous
andcanbeheterogeneousin
nature.
18/51
CharactericsofMulti-AgentSystems
Eachagenthasincomplete
informationorcapabilitiesfor
solvingtheproblemand,thus,has
alimitedviewpoint
2
Thereisnosystemglobalcontrol
3
Datais
4
Computationisasynchronous
1
Whatcantheydobetter?
TosolveProblemsthataretoolargefor
acentralizedagentwithlimited
distributed
Toreducetheriskoffailureofa
centralizedsystem
Disastermitigation/UrbanSearchAnd
Tokeeplegacysystemsinter-connectable
andinter-operational
Migrationofoutdated
Tosolveproblemsthatcannaturallybe
regardedassocietiesofautonomous
Air-trafficcontrol,Meetingscheduling,
19/51
20/51
ObjectOrientedProgrammingvsMulti-AgentSystems
Object-OrientedProgramming
Objectsarepassive,
i.e.anobjecthasno
controlovermethod
invocation
Objectsaredesigned
foracommongoal
Typicallyintegrated
intoasinglethread
ApplicationofMulti-AgentSystems(1/5):
ComputerGames
Multi-AgentSystems
Agentsare
autonomous,i.e.proactive
Agentscanhave
diverginggoals,e.g.
comingfromdifferent
organizations
Agentshaveown
threadofcontrol
RealTimeS trategy(e.g.S tarc raft,AgeofEmpires):TaskAssignmentandMulti-AgentPathPlanning
FirstPersonS hooter(e.g.HalfLife2,S plinterCell):CharacterInteractions,TeamFormation,LimitedSensing,PathPlanning,etc...
S imulations(e.g.TheS ims):CharacterInteractions,UtilityMaximization
Objectsdoitforfree;agentsdoitformoney.(Jenningsetal.1998)
21/51
ApplicationofMulti-AgentSystems(2/5):
LogisticandTransport
22/51
ApplicationofMulti-AgentSystems(3/5):
Industry
FactoryAndWarehouse
Management
Supplychain
management
B2B,Logistics
Taskassignment,coalitionformation,
pathplanning
ProjectKARIS:
Teamof100decentralized“elements”
toaccomplishautonomously
transportationtasks
coalitionformation
problem,standardized
communications,
auctions
Features:
Automaticloadandunloadatassembly
chains
Automaticbatteryrechargingviathe
ground
Mechanismtocouplewithstationsor
othervehicles
Airtrafficcontrol
distributedsensing,
auctions,...
Challenges:
Navigationandcoordinationof
decentralizedteams
23/51
24/51
ApplicationofMulti-AgentSystems(4/5):
SpaceExploration
ApplicationofMulti-AgentSystems(5/5):
Rescue
UrbanSearch
AndRescue
SpaceMissionswithmultiplerovers
distributedsensors
unmannedvehicles
firstresponder
management
Marsnetwork
Decentralizedsensing,
taskassignment,
coalitionformation,
pathplanning
Earthorbiters
Decentralizedsensing,taskassignment,coalitionformation...
Spaceshiprepair
...3Dpathplanning,andmanymorechallenges...
25/51
26/51
EmergencyServicesAssistance
Multi-AgentSystemsfor
EmergencyServicesAssistance
AdevastatingearthquakeofhighmagnitudeoccurredonDecember,26,
2004offthewestcoastofSumatra,Indonesia.TheresultingTsunami
killedthousandsofpeopleinsouthernIndia,Indonesia,Thailand,etc.
28/51
Fukushima
Tsunamisintheworld
11March2011,thenuclearfacility
atFukushimaisdestroyedbya
tsunami.
29/51
Thegolden72hours
30/51
MotivationforusingMulti-AgentSystems
Robotscanaccessplaceshumans
cannot
smallopeningsandconfinedspacesunderthe
floor
hazardousplaces
Mapping
Qualityofdisasterresponsestronglydependson
information,suchasmapswithvictimlocations
TomHaus(firemenat9/11):Weneedatracking
systemthattellsuswhereweare,wherewehave
been,andwherewehavetogoto
Efficientcoordinationofvictimsearch
Mixedinitiativeteamsofhumansandrobots
31/51
32/51
SearchandRelief
UASTechUAVPlatforms
Searchingforinjuredpeopleanddeliveringfood,
medicineandothersuppliesarehighlyprioritized
activitiesindisasterrelief.
LinkQuadweight~1kg,diameter~70cm
YamahaRMAXweight95kg,length3.6m
33/51
SelectedAutonomousFunctionalities
34/51
FindingInjuredPeople:Mission
Assumptions
Optimalflightaltitude:35-50m
Averageflightvelocity:5m/s
Humanbodysize:20-50
35/51
Context
11livebodies/2dummies
2UAVsfor
290x185m
36/51
FindingInjuredPeople:Results
DeliverFoodandSupplies:Mission
SaliencyMap
ThermalandColorImages
37/51
38/51
ACooperativeMissionScenario
Failtoattach
Ifthingscangowrongtheywillgowrong.
->needtomonitorexecution
39/51
40/51
SomefamousLandmarkProjects
RoboCup:
aLandmarkProject
forMulti-AgentSystems
TheApolloProgram:fromWrightFlyerin1903tothefirs tmanonthemoonin1969
ComputerChess:fromEniacin1946toDeepBluein1997
42/51
TheRoboCupProject
Thevision:By2050,buildateamoffullyautonomoushumanoids
whichwinsagainstthehumanworld-championundertheoffi cial
regulationofFIFA.
S inc e1997annualc ompetitionsandworkshops,sinc e2001
RoboCupResc ue
TheRoboCupProject
ComputerChessvsRoboCup
Feature
Chess
RoboCup
Environment Static
Dynamic
Aplatformforprojec t-orientededuc ationinsc ienc eand
AstandardproblemforAIand
Tec hnology
AlandmarkProjec t:c hallenginggoalandspill-overof
Worldaccessibility Completeinformation Incompleteinformation
Percepts
Execution
Non-symbolic
Symbolic
Turn-based Real-time
ActioneffectsDeterministic Stochastic
Agents
Central
Distributed
43/51
44/51
ThelandmarkofRoboCupRescue
RescueRobotCompetition
Bytheyear2050,enablelarge-scaleMulti-AgentSupportfor
disastermitigation
SensorNetworks
IntegrationofSensors
distributedinthecity
Simulatornetwork
e.g.FireGrid,RRSim
Humanrescuepersonnel
DigitallyEmpoweredbywearable
computers
SharedGISKnowledgebas ee.g.
GoogleMaps fors haringmis s ioncriticaldata
Roughterrainnavigation
EmergencyResponseCenter:
EfficientMASdecisionmaking
RobotTeamsReconnaissanceExplorationofinaccessibleplaces
45/51
RescueRobotCompetition
46/51
RescueVirtualCompetition
BasedontheUnrealgame
engine(UT2004,EpicGames)
RealisticmodelsofUSAR
environments,robots(Pioneer2DX,
SonyAIBO),andsensors(Laser
RangeFinder,ColorCamera,IMU,
WheelOdometry)
Multipleheterogeneousagentscan
beplacedinthesimulation
environment
Highfidelitysimulationofupto
12robots
AgentsconnectviaaTCP/IP
Simulationofwireless
Signsoflife:form,motion,heat,sound,CO2...
47/51
48/51
RescueAgentCompetion
RescueAgentCompetition
Thedomainmodelsalargeandcooperativemulti-agent
problem(#Agents>50)
Theenvironmentispartiallyobservable,agentshavetoact
rationallygiventhehistoryoftheirlocalpercepts
Thedomainisstochastic,effectsoffirefightingandrescue
mightvary
Theenvironmentissequential,i.e.continuouslyprogressing
Thedomainisdynamic,e.g.firesandcollapsingbuildings
Theworldisacontinuousandratherbig
Agentsareheterogeneoussincetheyhavedifferent
capabilities
Thedomainisdecentralizedduetoalimitedcommunication
bandwidth
Simulatorsfor
earthquake,fire,
civiliansandtraffic
Thetaskisto
developsoftware
agentswithdifferent
thatmakeroadspassable(police)
roles:
extinguishthefires(fire
rescueallcivilians(ambulances)
49/51
Summary
TolearnaboutMulti-AgentSystems
frombooksonlyisdifficult
Thereexistsnoultimatestrategyoralgorithm(maybe
inthefuture)
However,challengeswithindifferentdomainsarevery
similar
ForlearningaboutMulti-AgentSystemswehavetoget
intouchwiththem!
RoboCupRescueoffersarichsetof
ArtificialIntelligencerelated
problems:Letssolvethem!
51/51
50/51
Download