What is AI? Ar)ficial Intelligence. What is AI? What is AI? 9/4/09

Transcription

1 What is AI? Ar)ficial Intelligence CISC481/681 Lecture #1 Ben Cartere<e With material adapted from Prof. Marie desjardins (UMBC), Keith Decker, and Kathy McCoy. It is the science and engineering of making intelligent machines, especially intelligent computer programs. It is related to the similar task of using computers to understand human intelligence, but AI does not have to confine itself to methods that are biologically observable. John McCarthy, who coined the phrase OK, but what is intelligence? Copyright Ben Cartere<e 1 Copyright Ben Cartere<e 3 What is AI? Intelligence is the computa)onal part of the ability to achieve goals in the world. Varying kinds and degrees of intelligence occur in people, many animals and some machines. John McCarthy h<p://www formal.stanford.edu/jmc/wha)sai The exci)ng new effort to make computers think machines with minds, in the full and literal sense. (Haugeland, 1985) Like a human The art of crea)ng machines that perform func)ons that require intelligence when performed by people. (Kurzweil, 1990) What is AI? Think The study of the computa)ons that make it possible to perceive, reason, and act. (Winston, 1992) Well enough AI is concerned with intelligent behavior in ar)facts. (Nilsson, 1998) Act Copyright Ben Cartere<e 4 Copyright Ben Cartere<e 5 1

2 Ac)ng Like a Human Behaviorist approach Whatever the machine does, it should do it like a human Whether its thinking process is like a human is beside the point Turing Test: A person interviews an unseen en)ty to determine whether it is a human or a computer A computer that passes the test acts sufficiently human like in its understanding and use of language Thinking Like a Human The cogni)ve science approach Focus on reasoning process, not behavior or results Can biological brains teach us how computa)onal solu)ons should work? Can AI methods teach us anything about how biological brains func)on? The Chinese Room: A person in a room full of books takes ques)ons in Chinese from an interviewer The person doesn t understand Chinese at all The interview speaks Chinese fluently Looks up the ques)on in a book and returns the answer The Chinese Room passes the Turing Test, but it doesn t think like a human Copyright Ben Cartere<e 6 Copyright Ben Cartere<e 7 Ar)ficial Flight Thinking Ra)onally/Ideally/ Well Develop algorithms to compute formal models of Knowledge representa)on Reasoning Learning Memory Problem solving Focus on logical nota)on and inference All models are wrong, but some are useful George E. P. Box Copyright Ben Cartere<e 8 Copyright Ben Cartere<e 9 2

3 9/4/09 Ac)ng Ra)onally/Ideally/ Well So What is AI? For a given set of inputs, generate output that is good enough for the job at hand Heuris2cs: rules of thumb, strategies, tricks, simplifying devices, etc, that make a problem easier to solve Heuris)cs do not guarantee op)mal solu)ons; in fact, they do not guarantee any solu)on at all: all that can be said for a useful heuris1c is that it offers solu1ons which are good enough most of the 1me. Feigenbaum and Feldman, 1963, p. 6 One computer scien)st s view: Forget about intelligence we can t even define it! AI is a collec)on of: Problems that are not known to be have exact solu)ons or are not efficiently computable in prac)ce Techniques for a<acking such problems AI programs work right to the extent that they are useful to the people using them Firmly in the act well enough region 10 Other Ways of Organizing AI 11 What Can Heuris)c AI Do? Deduc)ve versus induc)ve approaches Play games Deduc)ve: start with a theory or model, confirm or falsify it with observa)ons Induc)ve: start with observa)ons, construct a theory or model that explains them Strong versus weak Strong : like a human; weak : not like a human Everything in this class would qualify as weak

4 9/4/09 Vision Natural language technologies 14 Scheduling and planning Drive a car 16 4

5 Make recommenda)ons Analyze medical imagery; make diagnoses Copyright Ben Cartere<e 18 Copyright Ben Cartere<e 19 Web Search Given a query, find pages that are most likely to be relevant. Rank them in decreasing order of likelihood of relevance. A search engine has to be intelligent Must be smart about narrowing down the search space Determining relevance requires an understanding of the user and the web page Humans cannot do what search engines do over the full web There is simply too much data Thinking and ac)ng like a human may be possible, but not desirable Search engines cannot do what humans can do with a single web page Current technology cannot understand a user or web page Thinking and ac)ng like a human not even possible Thinking ra)onally probably doesn t apply: web pages aren t necessarily logically coherent Brief History of AI AI is almost as old as computer science Which means it s only 50 or so years old Predates most of the big results in computa)onal complexity theory First a<empts were in the think/act like human region Quickly discovered that human brains are incredibly complex organs that are highly tuned for solving certain problems not easily duplicated What s easy for humans is not easy for computers (and vice versa) Today AI is one of the biggest disciplines in CS Many companies hiring AI focused programmers Google, Microsou, Yahoo, IBM, Oracle, Copyright Ben Cartere<e 20 Copyright Ben Cartere<e 22 5

6 AI is an Empirical Science AI problems ouen have either: Many possible correct solu)ons No efficiently computable exact solu)on Examples: Backgammon: Many possible strategies; few that consistently win Language understanding: I d like to thank my parents, Ayn Rand and God How to determine whether the solu)on provided by an AI is good? Empiricism: look at the evidence AI as an Empirical Science Which methods/techniques/representa)ons consistently give the best results? Best according to the person who s going to be using the system Evaluate the results over mul)ple trials Methods that are good for one problem or class of problems are not necessarily good for other problems Even problems that are very similar Copyright Ben Cartere<e 23 Copyright Ben Cartere<e 24 Course Goals About this course Primary goals: Understand AI problem solving techniques Determine when a problem requires AI methods to solve iden)fy problems as AI or not AI Analyze problems to determine which AI techniques are most appropriate Secondary goals: Empirically compare different AI solu)ons to a problem using sta)s)cal methods Copyright Ben Cartere<e 25 Copyright Ben Cartere<e 26 6

7 Topics Book: AI: A Modern Approach We ll cover a lot of material: Problem solving as search; search algorithms Knowledge representa)on; logic and inference Planning and scheduling Uncertainty and probability Learning Decision making Personnel Instructor: Ben Cartere<e carteret@cis.udel.edu Office: 401 Smith Hall Office hours: M 11:00 12:00, W 2:00 3:00 Appts by request (suggest a )me in your ) TA: Omer Arap Contact details TBA Copyright Ben Cartere<e 27 Copyright Ben Cartere<e 28 Course Web Page h<p://ir.cis.udel.edu/~carteret/cisc681 The best place to find up to date info on: Course descrip)on and policies Syllabus, schedule, lecture slides Homeworks, programming projects, readings Supplementary info Always check the web page before ing me or the TA with course ques)ons Homeworks Five wri<en assignments 30% of total grade (6% each) Turn in at the beginning of class Late policy: 20% off for each day auer due date If we get it during or auer class, it is one day late Five days late = 0% credit Ques)ons about problems or grading go to TA first Copyright Ben Cartere<e 29 Copyright Ben Cartere<e 30 7

8 Programming Assignments Four programming assignments 30% of total grade (7.5% each) Submit output and code to me and TA by midnight on the due date Late policy: 12:01am or later, it is one day late Ques)ons about assignment or grading go to TA first You may use any language that the TA or I understand: C, C++, Java, perl, python, R, Matlab python is recommended Exams Midterm and final exam Each 20% of total grade In class exams Final exam will be comprehensive Copyright Ben Cartere<e 31 Copyright Ben Cartere<e 32 Breakdown: Homeworks: 30% Programming: 30% Midterm: 20% Final: 20% Grades Grading scale: 90% for an A; 80% B; 70% C CISC481 vs. 681 Undergraduates graded on a separate curve, 5 points lower than graduate students 85% A; 70% B; 55% C Homeworks and exams may have problems for 681 students only Copyright Ben Cartere<e 33 Copyright Ben Cartere<e 34 8

9 Academic Integrity Thanks! Everything you hand in must be your own work No copying, no duplica)ng, no plagiarism If you discuss the problem with another student, write up your solu)ons later If you use online resources, write them in your own words and cite them Viola)ons will receive zero credit Ques)ons? Copyright Ben Cartere<e 35 Copyright Ben Cartere<e 36 9