cmpsci intelligence lecture september 8, 2011
TRANSCRIPT
![Page 1: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/1.jpg)
CMPSCI 383: Artificial IntelligenceLecture 2, September 8, 2011
Philip Thomas (TA)
![Page 2: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/2.jpg)
![Page 3: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/3.jpg)
Agent Function vs. Agent ProgramAgent Function vs. Agent Program
• Agent function: Maps any percept sequenceAgent function: Maps any percept sequence to an action
• Agent program: Internal implementation of• Agent program: Internal implementation of the agent functionT b l ?• Tabular?
![Page 4: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/4.jpg)
Vacuum‐Cleaner WorldVacuum Cleaner World
![Page 5: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/5.jpg)
Tabular Agent FunctionTabular Agent FunctionPercept Sequence Action
[A, Clean] Right
[A,Dirty] Suck
[B, Clean] Left
[B, Dirty] Suck
[A, Clean], [A, Clean] Right
[A, Clean],[A, Dirty] Suck
… …
[A, Clean], [A, Clean], [A, Clean] Right
… …
![Page 6: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/6.jpg)
Tabular Agent Program?Tabular Agent Program?
• Huge or infinite!Huge, or infinite!• How big would it be for chess?
E ti t d b f ibl t hi t i i– Estimated number of possible percept histories is 10^(10^50)Hardy, G. H. Ramanujan: Twelve Lectures on Subjects Suggested by His Life and Work, 3rd ed. New York: Chelsea, 1999.
• Can be represented more succinctly:– If dirty, suck, otherwise move to other locationy, ,
![Page 7: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/7.jpg)
Performance MeasurePerformance Measure
• “Pick up as much dirt as possible”p p– What might a rational agent do?
• “Keep the floor as clean as possible”
• Human arm movements“Mi i i i h d l i ”– “Minimize error in hand location”
– “Minimize error in joint angles”– “Minimize error in ^^^^^^^^, with penalty for using a lotMinimize error in , with penalty for using a lot of force”
– “Maximize the amount of time spent moving toward the goal while minimizing the time spent moving away”goal, while minimizing the time spent moving away
![Page 8: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/8.jpg)
ExamplesExamples
• Give me examples ofGive me examples of– Agent
• Sensors and PerceptsSensors and Percepts• Actuators and Actions
– EnvironmentEnvironment– Performance measure
![Page 9: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/9.jpg)
Rational AgentRational Agent
• For each possible percept sequence a rationalFor each possible percept sequence, a rational agent should select an action that is expected to maximize its performance measure givento maximize its performance measure, given the evidence provided by the percept sequence and whatever built‐in knowledgesequence and whatever built in knowledge the agent has.
![Page 10: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/10.jpg)
Rationality vs. PerfectionRationality vs. Perfection
• Crossing the street exampleCrossing the street example• Rationality: Maximize expected performance
f i i i l f• Perfection: Maximize actual performance
• Is it rational to cross a road without looking both ways?y
• Information gathering / exploration
![Page 11: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/11.jpg)
Black Sheep WallBlack Sheep Wall
Chess Poker
Checkers Blackjack
Tag Marco Polo
Fully Observable Partially Observable
Other Examples?
![Page 12: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/12.jpg)
Piano Poker
Running Tag
Balancing Marco Polo
Sudoku ChessSudoku Chess
Si l A M l i A
• Is driving single or multi‐agent?
Single Agent Multi Agent
![Page 13: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/13.jpg)
Driving (street) Pokerg ( )
Taxiing in Airplane Chess
Ninja Turtles arcade game Missile Guidance
Fire Truck Allocation Racing (car)Fire Truck Allocation Racing (car)
C i C i iCooperative Competitive
![Page 14: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/14.jpg)
Chess Fire Truck Allocation
Robot Control (ideally) Robot Control (in reality)
Vending Machine Monopoly
Poker D&DPoker D&D
D i i i E i S h i E i
• Stochastic = Random
Deterministic Environment Stochastic Environment
![Page 15: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/15.jpg)
Exploration vs. ExploitationExploration vs. Exploitation
• You’re new to a town and you find a way toYou re new to a town, and you find a way to work, it takes 10 minutes.
• You have 5 minutes before you have to be at• You have 5 minutes before you have to be at work for a meeting.D• Do you– 1) Take the route you know? (exploit)– 2) Try to take a shortcut that could waste time? (explore)
![Page 16: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/16.jpg)
Episodic vs. SequentialEpisodic vs. Sequential
• Episodic: Experience divided into atomicEpisodic: Experience divided into atomic episodes. In each episode the agent receives a percept and then performs a single actionpercept and then performs a single action.– The next episode does not depend on the actions taken in previous episodes.taken in previous episodes.
• Sequential: Current decision influences future decisionsdecisions.
![Page 17: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/17.jpg)
Chess Drivingg
Poker Robot Control
Games using the direction pad Games using the joystick
Di C iDiscrete Continuous
• State, time, action
![Page 18: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/18.jpg)
Reflex AgentReflex Agent
![Page 19: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/19.jpg)
Model‐Based Reflex AgentModel Based Reflex Agent
![Page 20: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/20.jpg)
Model‐Based Goal‐BasedModel Based Goal Based
![Page 21: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/21.jpg)
Model‐Based Utility‐BasedModel Based Utility Based
![Page 22: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/22.jpg)
Learning AgentLearning Agent
![Page 23: CMPSCI Intelligence Lecture September 8, 2011](https://reader034.vdocuments.mx/reader034/viewer/2022051603/62803a232f2a5751c5230faa/html5/thumbnails/23.jpg)
Atomic/Factored/StructuredAtomic/Factored/Structured