welcome!!! fifth americas agents school harvard university roger mailler sri international david...
TRANSCRIPT
Welcome!!!Welcome!!!
Fifth Americas Agents SchoolFifth Americas Agents School
Harvard UniversityHarvard University
Roger MaillerSRI International
David SarneHarvard University
Introduction to Agents and Introduction to Agents and Multi-agent SystemsMulti-agent Systems
What is an agent?What is an agent?
Gross DefinitionGross Definition– An agent is any objectAn agent is any object
SensesSensesDecidesDecidesActsActs
In the eye of the beholderIn the eye of the beholder– Thermostat can be considered an agentThermostat can be considered an agent– Not very intelligent!!!Not very intelligent!!!
Refining the DefinitionRefining the Definition
What makes an agent interesting is What makes an agent interesting is intelligenceintelligence
An intelligent agentAn intelligent agent– Senses its environmentSenses its environment– Makes decisions based onMakes decisions based on
What it senses (situated)What it senses (situated)Past experience (learning)Past experience (learning)To achieve a goal (goal driven)To achieve a goal (goal driven)
– Acts on those decisions without human Acts on those decisions without human intervention (autonomy)intervention (autonomy)
SensingSensing Sensing can be thought of as the act of collecting raw input Sensing can be thought of as the act of collecting raw input
and turning it into usable informationand turning it into usable information
The line between sensing and deciding is blurryThe line between sensing and deciding is blurry
Camera
Microphone
Laser
Text
ComputerVision
SpeechUnderstanding
Mapping
NLP
Images
Sounds
Points
Bytes
Decide Act
Error, Error, EverywhereError, Error, Everywhere
Sensing is an error-prone processSensing is an error-prone process– At the sensor levelAt the sensor level
CalibrationCalibration Insufficient ResolutionInsufficient Resolution Environmental InterferenceEnvironmental Interference
– Background noiseBackground noise– Multi-pathing Multi-pathing – Reflection and echoingReflection and echoing
– At the interpretation levelAt the interpretation level Incorrect translation induced by sensor level errorsIncorrect translation induced by sensor level errors Inaccurate and incomplete modelsInaccurate and incomplete models Translation ResolutionTranslation Resolution
Control
DecidingDeciding
Sense ActRules
Learning
Search
Control
Rule Based SystemsRule Based Systems Simplest way to act intelligently is a Rule-Based SystemSimplest way to act intelligently is a Rule-Based System Basically a large set of If-Then like rulesBasically a large set of If-Then like rules
If wind_speed > 30 thenIf wind_speed > 30 then increase thrust by 10%increase thrust by 10%
Highly engineeredHighly engineered Complicated to anticipate all conditions in the environmentComplicated to anticipate all conditions in the environment In some cases, it is impossible to have every solution pre-In some cases, it is impossible to have every solution pre-
programmedprogrammed
X Y
Z
If Y=blue and Z = greenX = red
If Y=blue and Z = redX = green
If Y=red and Z = blueX = green
And many more…
SearchSearch Useful when Useful when
– the number of possible solutions is largethe number of possible solutions is large– problem changes over timeproblem changes over time
Constraint satisfaction or Constraint satisfaction or optimization, scheduling, planningoptimization, scheduling, planning
X Y
Z
X
Y YY
Z Z Z
Y
Z Z Z
Search TechniquesSearch Techniques
Depends largely on the problemDepends largely on the problem Some methodsSome methods
– BacktrackingBacktracking– OR techniques like SimplexOR techniques like Simplex– Flow TechniquesFlow Techniques– Local SearchLocal Search
Hill ClimbingHill ClimbingSimulated AnnealingSimulated Annealing
LearningLearning Simplest form of learning is lookup tablesSimplest form of learning is lookup tables
– Store past experience and outcomeStore past experience and outcome– Assume that a repeated experience has the Assume that a repeated experience has the
same outcomesame outcome ProsPros
– Easy to implementEasy to implement– UnbiasedUnbiased
ConsCons– Memory intensiveMemory intensive– Only can predict something that you’ve Only can predict something that you’ve
previously seenpreviously seen– Not robust to ambiguityNot robust to ambiguity
Other learning techniquesOther learning techniques Decision TreesDecision Trees
– Pros: Easy to understand and implementPros: Easy to understand and implement– Cons: Biased and tend to over-fit dataCons: Biased and tend to over-fit data
Neural NetworksNeural Networks– Pros: Can fit almost any functionPros: Can fit almost any function– Cons: Slow to train, tend to over-fit, impossible to figure out Cons: Slow to train, tend to over-fit, impossible to figure out
what they’ve learned what they’ve learned Reinforcement LearningReinforcement Learning
– Pros: Guaranteed to find an optimal policy in the limitPros: Guaranteed to find an optimal policy in the limit– Cons: Slow to train, need a fairly robust model of the Cons: Slow to train, need a fairly robust model of the
environment (states and actions)environment (states and actions) Genetic AlgorithmsGenetic Algorithms
– Pros: Great for learning parametersPros: Great for learning parameters– Cons: Hard to setup, very slowCons: Hard to setup, very slow
Inductive Logic ProgrammingInductive Logic Programming– Pros: Can understand what’s learned, knowledge is consistentPros: Can understand what’s learned, knowledge is consistent– Cons: Has a hard time dealing with inconsistent knowledge Cons: Has a hard time dealing with inconsistent knowledge
ControlControl
Levels of controlLevels of control– Single execution threadSingle execution thread– Finite State MachineFinite State Machine– ThreadsThreads– ProcessProcess
Meta-level reasoningMeta-level reasoning– How long should I take to reason about this How long should I take to reason about this
input?input?– How much time should I spend searching How much time should I spend searching
versus learning?versus learning?
ActingActing
Sense Decide
Effectors
Graphics
Speech
Text
More ErrorsMore Errors
Inaccurate modelsInaccurate models– KinematicsKinematics– Grammar and languageGrammar and language– PronunciationPronunciation
ResolutionResolution– Details matterDetails matter
CoordinationCoordination In many environments, agents’ actions effect one In many environments, agents’ actions effect one
anotheranother For example, if I decide to drive through an For example, if I decide to drive through an
intersection, my actions effect any other agent intersection, my actions effect any other agent trying to drive through at the same timetrying to drive through at the same time
Reasons to coordinateReasons to coordinate– Shared resource – two cars in the same intersectionShared resource – two cars in the same intersection– Shared goal requiring multiple agents – building a houseShared goal requiring multiple agents – building a house– Overlapping goals, my action helps you out without Overlapping goals, my action helps you out without
costing me anythingcosting me anything
Forms of CoordinationForms of Coordination Implicit (without communication)Implicit (without communication)
– Social policies, norms, rulesSocial policies, norms, rules Stay on the right hand side of the roadStay on the right hand side of the road Steam (Tambe et al.)Steam (Tambe et al.)
– Observation (communicating through the environment)Observation (communicating through the environment) I start driving forward so you better stopI start driving forward so you better stop Swarms (Bonabeau, Eberhart, Shi, Kennedy )Swarms (Bonabeau, Eberhart, Shi, Kennedy )
Explicit (with communication)Explicit (with communication)– CompetitiveCompetitive
Auctions (Wellman, Sandholm)Auctions (Wellman, Sandholm)– CooperativeCooperative
Negotiation (Lesser, Durfee)Negotiation (Lesser, Durfee) Mediation (Cammerata ,Shen, Mailler)Mediation (Cammerata ,Shen, Mailler)
Implicit CoordinationImplicit Coordination Social norms and policies: simply adding Social norms and policies: simply adding
rules to follow in social settingsrules to follow in social settings– Pros: No need to communicate, fast to Pros: No need to communicate, fast to
compute a solutioncompute a solution– Cons: Like rule-based system, require a lot of Cons: Like rule-based system, require a lot of
pre-planning and design, can interfere with pre-planning and design, can interfere with internal rulesinternal rules
Swarms (everyone do there own thing)Swarms (everyone do there own thing)– Pros: Work well when individual optimality is Pros: Work well when individual optimality is
not importantnot important– Cons: Hard to create micro behaviors that Cons: Hard to create micro behaviors that
create a macro solution create a macro solution
Market-based MechanismsMarket-based Mechanisms Contract Net (Smith) – Announce contract, Contract Net (Smith) – Announce contract,
agents bid, contract is awardedagents bid, contract is awarded AuctionsAuctions
– English, Dutch, Vickery, Double, CombinatorialEnglish, Dutch, Vickery, Double, Combinatorial When they workWhen they work
– Utility of the bidding agents is independent Utility of the bidding agents is independent from one another (utility is not super-additive)from one another (utility is not super-additive)
– Market is designed to prevent collusion (often Market is designed to prevent collusion (often hard to do)hard to do)
– Auctioneer is trustedAuctioneer is trusted– Agents are not willing to reveal a lot of Agents are not willing to reveal a lot of
informationinformation
NegotiationNegotiation
Semi-cooperativeSemi-cooperative Willing to exchange some information in Willing to exchange some information in
order to resolve an impasseorder to resolve an impasse Usually involves multiple steps to Usually involves multiple steps to
converge on a solutionconverge on a solution Works well when Works well when
– dependencies are localized (no complex multi-dependencies are localized (no complex multi-linking)linking)
– Agents are willing to concede local utility for Agents are willing to concede local utility for the greater goodthe greater good
Cooperative MediationCooperative Mediation
Cooperative methodCooperative method Willing to eventually reveal all Willing to eventually reveal all
information to converge on a solutioninformation to converge on a solution Works well whenWorks well when
– Complex interdependencies between Complex interdependencies between agentsagents
– Agents are willing to exchange local Agents are willing to exchange local utility for global utilityutility for global utility
ConclusionsConclusions
An intelligent agent is a goal-directed, An intelligent agent is a goal-directed, autonomous object which is capable of autonomous object which is capable of sensing, deciding, and acting.sensing, deciding, and acting.
Agents represent the pinnacle of current Agents represent the pinnacle of current and past AI researchand past AI research– LearningLearning– SearchSearch– CoordinationCoordination– SensingSensing– EffectingEffecting
ScheduleSchedule Friday July 14thFriday July 14th
– 11:00-12:30 11:00-12:30 Computational Mechanism DesignComputational Mechanism Design David David ParkesParkes
– 2:00-3:30 2:00-3:30 Autonomous Bidding Agents: Lessons from the Autonomous Bidding Agents: Lessons from the Trading Agent CompetitionTrading Agent Competition Amy Amy GeenwaldGeenwald
– 4:00-5:30 4:00-5:30 Bio-inspired Multi-Agent SystemsBio-inspired Multi-Agent Systems RadhikaRadhika NagpalNagpal
– 6:00 -8:00 6:00 -8:00 Reception and Poster SessionReception and Poster Session Saturday July 15thSaturday July 15th
– 9:00-10:30 9:00-10:30 Graphical Models for Decision-Making in MASGraphical Models for Decision-Making in MAS Gal Gal KobiKobi
– 11:00-12:30 11:00-12:30 Ant RobotsAnt Robots Sven Koenig Sven Koenig– 2:00-3:30 2:00-3:30 Key Issues in Multi-Agent CoordinationKey Issues in Multi-Agent Coordination Victor Victor
LesserLesser– 4:00-5:30 4:00-5:30 Understanding Issues in Multi-Agent System Understanding Issues in Multi-Agent System
Design Through the Blackboard System Model.Design Through the Blackboard System Model. Zack Zack RubinsteinRubinstein