CNS 4470CNS 4470
Artificial IntelligenceArtificial Intelligence
What is AI?What is AI?
• No really what is it?No really what is it?
Acting Humanly: Acting Humanly: The Turing TestThe Turing Test
• Allen Turing (1950) predicted:Allen Turing (1950) predicted:– By 2000, a machine might have a 30% By 2000, a machine might have a 30%
chance of fooling a lay person for 5 chance of fooling a lay person for 5 minutesminutes
– He was nearly right.He was nearly right.•True if you restrict the domain to something True if you restrict the domain to something
like Law.like Law.
– Suggested major components of AI:Suggested major components of AI:•knowledge, reasoning, Language knowledge, reasoning, Language
understanding and learning.understanding and learning.
Thinking Humanly:Thinking Humanly:Cognitive ScienceCognitive Science
• Requires scientific theories of internal Requires scientific theories of internal activities of the brainactivities of the brain– What level of abstraction?What level of abstraction?
•Knowledge vs. CircuitsKnowledge vs. Circuits
– How to validate: requiresHow to validate: requires•Predicting and testing behavior of human Predicting and testing behavior of human
subjectssubjects
•Direct identification from neurological dataDirect identification from neurological data
• Now separate sciences from AINow separate sciences from AI
Thinking rationally:Thinking rationally:Laws of ThoughtLaws of Thought
• Greek philosophy to present day math and Greek philosophy to present day math and philosophy have prescribed “right-philosophy have prescribed “right-thinking”thinking”
• But…But…– Not all intelligent behavior requires logical Not all intelligent behavior requires logical
deliberationdeliberation– How do I really know what I should be thinking How do I really know what I should be thinking
Right Now!Right Now!
Acting RationallyActing Rationally
• Rational behavior: Do the right thingRational behavior: Do the right thing
• The right thing:The right thing:– Maximize goal achievementMaximize goal achievement– Given available informationGiven available information
• Not necessarily involve “thinking”Not necessarily involve “thinking”– ReflexReflex– Detailed PlanningDetailed Planning
HistoryHistory
• 1943-19551943-1955– McCulloch and PittsMcCulloch and Pitts– Donald HebbDonald Hebb– Hebbian LearningHebbian Learning
– Refuted by Marvin MinskyRefuted by Marvin Minsky
HistoryHistory
• Birth – 1956Birth – 1956– John McCarthy workshop at DartmouthJohn McCarthy workshop at Dartmouth– Allen NewellAllen Newell– Herb SimonHerb Simon
• Logic TheoristLogic Theorist• Solved Mind – body problemSolved Mind – body problem
– Showed that AI was only field to attempt to Showed that AI was only field to attempt to build machines that would function build machines that would function autonomously in complex, changing autonomously in complex, changing environmentsenvironments
HistoryHistory
• 1952 – 19691952 – 1969– ““Look mom no hands”Look mom no hands”
– General Problem SolverGeneral Problem Solver– LispLisp– AnalogyAnalogy
•geometry solving programgeometry solving program
– PerceptronsPerceptrons
HistoryHistory
• Dose of realityDose of reality– 1966 – 19731966 – 1973– No world Chess championNo world Chess champion– No translator between Russian and EnglishNo translator between Russian and English
•Spirit is willing but the flesh is weakSpirit is willing but the flesh is weak•the vodka is good but the meat is rottenthe vodka is good but the meat is rotten
– Summarized as Failure to come to grips Summarized as Failure to come to grips with the “combinatorial explosion”with the “combinatorial explosion”
HistoryHistory
• 1969-19791969-1979
• Expert SystemsExpert Systems– DendralDendral
•Reasoning about chemistryReasoning about chemistry
– MycinMycin•Diagnose blood infectionsDiagnose blood infections
HistoryHistory
• 1987 - Present1987 - Present• AI becomes ScienceAI becomes Science
– AI no longer separated from the rest of AI no longer separated from the rest of computer sciencecomputer science
– Ex. Machine learning not separated from Ex. Machine learning not separated from information theoryinformation theory
• ExamplesExamples– Backprop Neural NetworksBackprop Neural Networks– Hidden Markov ModelsHidden Markov Models– Bayesian NetworksBayesian Networks
Rational AgentsRational Agents
• AgentAgent– PerceivesPerceives– ActsActs
• FunctionFunction– Precepts --> ActionsPrecepts --> Actions
• Seek agent with best performanceSeek agent with best performance• Or given computational constraintsOr given computational constraints
– Best program, given machine resourcesBest program, given machine resources
Rational AgentsRational Agents
• Not alwaysNot always– OmniscientOmniscient– ClairvoyantClairvoyant– SuccessfulSuccessful
• Just Act Reasonably given the Just Act Reasonably given the information at handinformation at hand
Computer-HumanComputer-HumanComparisonComparison
1CPU1CPU
10^10 bits RAM10^10 bits RAM
10^12 bits disk10^12 bits disk
10^9 cycles/sec10^9 cycles/sec
10^10 bits/sec10^10 bits/sec
10^9 updates/sec10^9 updates/sec
10^11 neurons10^11 neurons
10^11 neurons10^11 neurons
10^14 synapses10^14 synapses
10^ 3 cycles/sec10^ 3 cycles/sec
10^14 bits/sec10^14 bits/sec
10^14 updates/sec10^14 updates/sec
State of the ArtState of the Art
• Autonomous Planning and SchedulingAutonomous Planning and Scheduling– Remote Agent: Actual Scheduling of operations Remote Agent: Actual Scheduling of operations
of a NASA Spacecraftof a NASA Spacecraft• Game playingGame playing
– IBM’s Deep Blue defeated the world championIBM’s Deep Blue defeated the world champion• Autonomous controlAutonomous control
– ALVINN: Controlled a minivan across the US ALVINN: Controlled a minivan across the US 98% of the time.98% of the time.
• DiagnosisDiagnosis– Expert physician diagnosis: Medical expert Expert physician diagnosis: Medical expert
scoffs then later agrees with the diagnosisscoffs then later agrees with the diagnosis
State of the ArtState of the Art
• Logistics PlanningLogistics Planning– DART: Dynamic Analysis and Re-planning Tool: DART: Dynamic Analysis and Re-planning Tool:
50,000 vehicles, cargo and people50,000 vehicles, cargo and people
• RoboticsRobotics– Robotic micro-surgery assistants: three-Robotic micro-surgery assistants: three-
dimensional model of internal anatomy & guided dimensional model of internal anatomy & guided insertion of artificial hipinsertion of artificial hip
• Language Understanding and Problem Language Understanding and Problem SolvingSolving– Proverb: Solves crossword puzzles better than Proverb: Solves crossword puzzles better than
most humansmost humans
Setting for Intelligent Setting for Intelligent DesignDesign
• P.A.G.E.P.A.G.E.
• PPerceptsercepts
• AActionsctions
• GGoalsoals
• EEnvironmentnvironment
Types of AgentsTypes of Agents
• Reflex AgentsReflex Agents– SensorsSensors– Rules -> ActionsRules -> Actions
• Reflex Agents with StateReflex Agents with State– Add the ability to know how actions Add the ability to know how actions effect the current effect the current
statestate
• Goal-Based AgentsGoal-Based Agents– Add the Ability to know how actions effect Add the Ability to know how actions effect future statesfuture states
• Utility-Based AgentsUtility-Based Agents– Add the Ability to know how Add the Ability to know how happyhappy it will be under it will be under
various goal conditionsvarious goal conditions
Performance MeasurePerformance Measure
• Objective measure of Agent’s Objective measure of Agent’s performanceperformance– Specified by the DesignerSpecified by the Designer– Connected to the Agent’s GoalConnected to the Agent’s Goal
• Intelligent AgentIntelligent Agent– Maximizes the Performance MeasureMaximizes the Performance Measure
Given available resourcesGiven available resources
EnvironmentsEnvironments
• Fully observable vs. partially Fully observable vs. partially observableobservable
• Deterministic vs. StochasticDeterministic vs. Stochastic
• Episodic vs. sequentialEpisodic vs. sequential
• Static vs. DynamicStatic vs. Dynamic
• Discrete vs. ContinuousDiscrete vs. Continuous
• Single agent vs. multi-agent.Single agent vs. multi-agent.