CS 4100 Artificial al Intelligence or: Jan-Willem van de Meent In Instructor ite: https://course.ccs.neu.edu/cs4100f19 Websit At Attribution : many of these slides are modified versions of those distributed with the UC Berkeley CS188 materials Thanks to John DeNero, Dan Klein, and Peter Abbeel.
Today ay’s plan an • What is AI? Plus a brief history • What this course is about and some logistics
Defining Artificial al Intelligence • Precise definitions are surprisingly elusive • Informally ally: The discipline of creating intelligent algorithms • Here we’ve just offloaded the complexity into the term intelligent • In In press: Any algorithm that makes predictions • AI often means the same thing as machine learning • Most machine learning is essentially computational statistics
The Brai ain as as Inspirat ation for AI Humans are in some case very good at making • rational decisions, but certainly not perfect Brains are very hard to reverse engineer! • “ Brains are to intelligence as wings are to flight ” • Lessons learned from the brain: memory and • simulation are key to decision making
Thi This Cour urse: e: Rat ational al approaches to thinking and acting
Rat ational al decisions Here we use the term rat ational al in a specific, technical manner • Rat ational al: maximally achieving pre-defined goals • Rationality only concerns what de decisions are made (not the thought processes underpinning them) • Goal als are expressed in terms of the ut utility of ou outc tcom omes • Being rat ational al means max aximizing your expected utility We may think of this view of AI as Computational Rationality
The imperative: maximize expected utility
To Topi pics cs in n Thi This Co Cour urse Par art I: Mak aking Decisions (Acting Rationally) Par art II: Reas asoning under Uncertai ainty (Thinking Rationally) Par art III: Ethics of AI Thr Throug ugho hout: ut: Applications and emerging research
Age Agent-bas ased Approac aches to Mak aking Decisions • We will be concerned with designing ag agents that • operate in some env environm nment ent (very broadly interpreted) • maximize some notion of expected uti utility ty • Typically there are pe ions available and ac actions we can take percept ptio ates (e.g., locations) • The agent operates in stat • The environment might be stat atic ( constant ) or dynam amic ( changing ) • It also might be fu lly or only par ally observable fully artial
Vac acuum-clean aner world The agent might be a Roomba
Vac acuum-clean aner world The agent might be a Roomba We have: some perceptions via sensors •
Vac acuum-clean aner world The agent might be a Roomba We have: some perceptions via sensors • actions we can take (suck, move right, …) •
Vac acuum-clean aner world The agent might be a Roomba We have: some perceptions via sensors • actions we can take (suck, move right, …) • • utility we’d like to maximize (some combination of ”don’t break” and ”clean all floors”)
Vac acuum-clean aner world The agent might be a Roomba We have: some perceptions via sensors • actions we can take (suck, move right, …) • • utility we’d like to maximize (some combination of ”don’t break” and ”clean all floors”) • states where are we on the floor?
Rat ational al vac acuum clean aner?
Rat ational al vac acuum clean aner Here actions depend deterministically on states; i.e., this is a look up table. We will call such agents reflex ag agents .
Rat ational al vac acuum clean aner Here actions depend deterministically on states; i.e., this is a look up table. We will call such agents reflex ag agents . Ques Question: n: How would we evaluate such an agent?
Defining a a Good Notion of Utility is Har ard • One meas asure of performan ance: Amount of dirt cleaned in an eight-hour shift. • Pr Problem: The agent can maximize this performance by cleaning the floor, then dumping out all the dirt, and then cleaning it again. • Better meas asure of performan ance: Amount of time that floor is clean. • Ru Rule of Thumb (from Textbook) k): It is better to design utility according to outcomes, than according to how we think the agent should behave.
A A brie ief sele lectiv tive his istor tory of of AI AI
A brief an and selective history of AI 1940 1940-1950: 1950: Early y da days ys • - 1943: McCulloch & Pitts: Boolean circuit model of brain - 1950: Turing's “Computing Machinery and Intelligence”
Min Minsk sky Built the first Neural Network computer in 1950
Min Minsk sky Built the first Neural Network computer in 1950 The same year that Turing proposed his test
The Turi The Turing ng Test Test Image credit: http://turing100.blogspot.com/2012/05/one-month-to-biggest-turing-test.html
https://xkcd.com/329/
A brief an and selective history of AI 1940 1940-1950: 1950: Early y da days ys • - 1943: McCulloch & Pitts: Boolean circuit model of brain - 1950: Turing's “Computing Machinery and Intelligence” 1950 1950—70: 70: Exc xcitement – Lo Look, Ma, a, no han ands! • - 1950s: Early AI programs, including Samuel's checkers program, Newell & Simon's Logic Theorist, Gelernter's Geometry Engine - 1956: Dartmouth meeting: “Artificial Intelligence” adopted - 1965: Robinson's complete algorithm for logical reasoning
Initial al Successes: Checkers
Bold clai aims an and extreme optimism “ machines will be capable, within twenty years, of doing any work a man can do” - Herbert Simon, in 1965
Initial al Successes: Toy Worlds an and Sear arch
A brief an and selective history of AI 1940 1940-1950: 1950: Early y da days ys • - 1943: McCulloch & Pitts: Boolean circuit model of brain - 1950: Turing's “Computing Machinery and Intelligence” 1950 1950—70: 70: Exc xcitement – Lo Look, Ma, a, no han ands! • - 1950s: Early AI programs, including Samuel's checkers program, Newell & Simon's Logic Theorist, Gelernter's Geometry Engine - 1956: Dartmouth meeting: “Artificial Intelligence” adopted - 1965: Robinson's complete algorithm for logical reasoning 1970—90: 1970 90: Knowledg dge-ba based d appr pproaches • - 1969—79: Early development of knowledge-based systems - 1980—88: Expert systems industry booms - 1988—93: Expert systems industry busts: “AI Winter”
A brief an and selective history of AI 1940 1940-1950: 1950: Early y da days ys • - 1943: McCulloch & Pitts: Boolean circuit model of brain - 1950: Turing's “Computing Machinery and Intelligence” 1950—70: 1950 70: Exc xcitement – Lo Look, Ma, a, no han ands! • - 1950s: Early AI programs, including Samuel's checkers program, Newell & Simon's Logic Theorist, Gelernter's Geometry Engine - 1956: Dartmouth meeting: “Artificial Intelligence” adopted - 1965: Robinson's complete algorithm for logical reasoning 1970 1970—90: 90: Knowledg dge-ba based d appr pproaches • - 1969—79: Early development of knowledge-based systems - 1980—88: Expert systems industry booms - 1988—93: Expert systems industry busts: “AI Winter” 1990 1990—2012: 2012: Statistical appr pproaches • - Resurgence of probability, focus on uncertainty - General increase in technical depth - Agents and learning systems… “AI Spring”?
A brief an and selective history of AI 1940 1940-1950: 1950: Early y da days ys • - 1943: McCulloch & Pitts: Boolean circuit model of brain - 1950: Turing's “Computing Machinery and Intelligence” 1950 1950—70: 70: Exc xcitement – Lo Look, Ma, a, no han ands! • - 1950s: Early AI programs, including Samuel's checkers program, Newell & Simon's Logic Theorist, Gelernter's Geometry Engine - 1956: Dartmouth meeting: “Artificial Intelligence” adopted - 1965: Robinson's complete algorithm for logical reasoning 1970 1970—90: 90: Knowledg dge-ba based d appr pproaches • - 1969—79: Early development of knowledge-based systems - 1980—88: Expert systems industry booms - 1988—93: Expert systems industry busts: “AI Winter” 1990 1990—2012: 2012: Statistical appr pproaches • - Resurgence of probability, focus on uncertainty - General increase in technical depth - Agents and learning systems… “AI Spring”? 2012 2012—: E : Excitement – Lo Look, Ma, a, no han ands ag agai ain? • - Big data, big compute, neural networks - Some re-unification of sub-fields - AI used in many industries
AI / ML L is star arting to be everywhere • Advertising • Search engines • Route planning • Spam / fraud detection • Automated help desks • Product recommendations
Recommend
More recommend
