lect9_artificialintelligence - Introductions to Modern...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: Introductions to Modern Computer Science Artificial Intelligence (AI) Professor Adnan Darwiche [email protected] http://cs.ucla.edu/~darwiche Fall 2007 CS 1: What is AI? A field of computer science concerned with understanding and building intelligent agents that can: Perceive Understand Predict Manipulate Learn This is very hard, yet possible, quite useful and a lot of fun! Hardware versus software (internet) agents Monolithic approach unnecessary for usefulness What is AI? Broad goals of AI are shared with other disciplines, including philosophy, cognitive science (psychology) and neuroscience. Some key distinctions: Understand versus build Rational versus human-like Specific goals of AI are shared with other disciplines, including linguistics (perception), statistics (learning), and mathematical logic (understanding and prediction). What is AI? AI is very old: Philosophers have been thinking about reasoning, perception and other mental abilities for over 2000 years AI is very recent: "Artificial Intelligence" coined in 1956 Computers as tools for testing theories of intelligence AI is popular: Largest number of applications to graduate program at UCLA Often cited as "field I would most like to be in" along with modern genetics The Turing Test Proposed by Alan Turing in 1950 as an operational test of intelligence: fool a human interrogator into believing the agent is a human. To pass the test, one needs: Natural language processing Knowledge representation Automated reasoning Machine learning Avoided physical interaction between interrogator and agent (physical simulation is unnecessary for intelligence), excluding need for: Computer vision Robotics See lecture by Prof. Demetri Terzopoulos The Intelligent Agent Architecture Perception: written language speech vision touch mouse click ... Knowledge Learning Embedding Environment Action: written language speech movement information display ... Reasoning Knowledge Representation How do we represent knowledge? Factual knowledge (facts) Uncertain knowledge (beliefs) What knowledge is relevant? How do we acquire knowledge? From experts By conversion from other forms of knowledge By learning from experience KR: How do we represent knowledge? Facts Knowledge: Beliefs Knowledge: There are four people: Roberta, Thelma, Steve, A few weeks after inseminating a cow, we have three and Pete. possible tests to confirm pregnancy. The first is a Among them, they hold eight different jobs. scanning test which has a false positive of 1% and a false Each holds exactly two jobs. negative of 10%. The second is a blood test, which detects The jobs are chef, guard, nurse, clerk, police officer progesterone with a false positive of 10% and a false First order logic (Implicit versus explicit knowledge): negative of 30%. The third test is a urine test, which also (gender not implied), teacher, actor, and boxer. The job of nurse is held by a male. then HASAJOB(x,clerk) detects progesterone with a false positive of 10% and a If HUSBAND(x,jobholder(chef)) The husband of the chef is the clerk. false negative of 20%. The probability of a detectable FEMALE(jobholder(chef)) Roberta is not a boxer. progesterone level is 90% given pregnancy, and 1% given Pete HUSBAND(x,y), then MALE(x)grade. If has no education past the ninth and FEMALE(Y) no pregnancy. The probability that insemination will Roberta, the chef, and the police officer went impregnate a cow is 87%. If HASAJOB(x,nurse) then GREATERTHAN(education(x),9) golfing together. Question: What's the belief in pregnancy given a postive ... Question: Who holds which jobs? urine test but a negative blood test? ... Representations: Propositional logic First order logic ... Representations: Belief networks Fuzzy logic ... Reasoning How can we formalize the reasoning process? Deduction: What is implied by a knowledge base? Belief revision: What beliefs to give up in case of a contradiction? Causality: What is the cause of an event? How can we reason efficiently? Time Space Footprint Reasoning: Formalizing Belief Revision Oscar used to believe that he had given Victoria a gold ring at their wedding. He had bought their two wedding rings at a jeweler's shop in Casablanca. He thought it was a bargain. The merchant had claimed that the rings were made of 24 carat gold. They certainly looked like gold, but to be on the safe side Oscar had taken the rings to the jeweler next door who has testified to their gold content. However, some time after the wedding, Oscar was repairing his boat and he noticed that the sulfuric acid he was using stained his ring. He remembered from his school chemistry that the only acid that affected gold was aqua regia. Somewhat surprised, he verified that Victoria's ring was also sustained by the acid. So Oscar had to revise his beliefs because they entailed an inconsistency. He could not deny that the rings were stained. He toyed with the idea that, by accident, he had bought aqua regia rather than sulfuric acid, but he soon gave up this idea. So, because he had greater confidence in what he was taught in chemistry than in his own smartness, Oscar somewhat downheartedly accepted that the rings were not made of gold after all. Consequently, he was convinced that both jewelers had been lying. He also came to believe that they were in collusion with each other, although he was not completely certain of this. Reasoning: Formalizing Causality When do we say A caused B? Needed for explanation Allow us to predict the future Suggest ways to control future events Moral responsibility Legal liability It is not simply sufficiency, nor necessity... Billy and Suzy both throw rocks at a bottle. Suzy's arm is better than Billy's, so her rock gets to the bottle first and shatters it. Billy's throw was perfectly accurate, so his rock would have shattered the bottle had Suzy's missed. Did Billy's throw cause the bottle to shatter? What if both rocks hit the bottle simultaneously. Gene 1 Gene 2 Gene 3 Applications of KR&R Medical diagnosis Credit card fraud Theorem proving (Mathematics) Formal verification Cognitive tasks: planning, explanation, etc The Intelligent Agent Architecture Perception: written language speech vision touch mouse click ... Knowledge Learning Embedding Environment Action: written language speech movement information display ... Reasoning Natural Language Understanding Understand natural language: List all employees that have been working here for more than two years and have not gotten a raise since then. Generate natural language: There are actually only two such employees, do you just need their names? Text summarization: Here's a 100 word abstract of the article... Machine translation Convert text from one language into another. Natural Language Understanding Why is it hard? Context and ambiguity. Syntactic ambiguity: "They are cooking apples" Semantic ambiguity: "She ran to the bank" Pragmatic ambiguity: "Can you open the door?" Two key approaches: Classical: Provide the system with rich knowledge of the world, perform thorough syntactic, semantic and pragmatic analysis, disambiguate using knowledge and reasoning. Modern: Rely on corpus and disambiguate using machine learning. Ambiguity and Context They are cooking apples sentence -> noun phrase , verb phrase noun phrase -> noun noun phrase -> determiner , noun verb phrase -> verb , noun phrase Natural Language Understanding: Web Translation Natural Language Understanding: Eliza the Chatbot! Try it: http://www-ai.ijs.si/eliza-cgi-bin/eliza_script Adnan: I want to give an introductory lecture on AI. Eliza: Why do you want to give an introductory lecture on AI? Adnan: Our department chair asked me to. Eliza: Can you elaborate on that? Adnan: He says multiple faculty members will contribute to this class. Eliza: Please go on. Adnan: I don't think you understood what I just said. Eliza: Why do you not think I understood what you just said? Adnan: Because you seem to be rambling. Eliza: Oh, i seem to be rambling. Adnan: Absolutely. Eliza: Please go on. Adnan: I give up! Eliza: Can you elaborate on that? Adnan: You just proved my claim. Eliza: Why do you say your claim? Adnan: Because I made it. Eliza: Do any other reasons not come to mind? Adnan: I think the class have gotten the point about you by now. Eliza: Do you wish to think the class have gotten the point about me? Adnan: Yes. Eliza: Are you sure? .... The Intelligent Agent Architecture Perception: written language speech vision touch mouse click ... Knowledge Learning Embedding Environment Action: written language speech movement information display ... Reasoning Machine Learning Using experiences and observations to improve future performance (actions): What aspect of performance to be improved? Irrelevant aspects of the world, how the world evolves, what are desirable/undesirable situations. What feedback is available? Supervised, Unsupervised, Reinforcement learning. How to represent the output of a learning process? Logical knowledge, Probabilistic knowledge, Neural networks. Supervised: Give observations and actions they should lead to Unsupervised: Give observations only (find patterns ) Reinforcement: Give positive/negative feedback on actions Supervised Learning: Character Recognition Neural Networks Unsupervised Learning: Recommender Systems E-commerce Recommend products based on previous purchases or click-stream behavior Ex: Amazon.com Information sites Rate items based on previous user ratings Ex: MovieLens, Jester Unsupervised Learning: Recommender Systems Offline Learning John Sam Cindy 5 3 4 - 3 1 5 2 5 - Learned Knowledge Reasoning Bob 5 1 3.5 1.7 Planning Finding a sequence of actions that will achieve a goal Input: Actions (preconditions, effects) Initial, Goal states Knowledge of world (physics) Output: Plan: Conditional (contingency) Total/Partial Sensorless (conformant) Autonomous Sciencecraft Experiment: uses on-board science analysis and re-planning to radically increase science return by enabling intelligent downlink selection and autonomous retargeting. Multi-Rover Integrated Science Understanding System: develops architectures and technologies for command and control of multi-rover groups for planetary exploration. NASA: Autonomous Sciencecraft Experiment Use onboard decision-making to detect, analyze, and respond to science events, and to downlink only the highest value science data. Radically increase science return by enabling intelligent downlink selection and autonomous retargeting. AI Technology Used: Image analysis Planning, scheduling and recovery Mission status updates: http://www-aig.jpl.nasa.gov/public/planning Robotics Robots: Physical agents that perform tasks by manipulating the world. Equipped with: Effectors: legs, wheels, joints, grippers Sensors: cameras, ultrasound, gyroscopes, accelerometers. Common categories of robots: Manipulators (robot arms): factory assembly lines, international space station. Mobile robots: unmanned land/air/water vehicles, planetary rovers Mobile robots with manipulators: humanoid (mimic human torso) Robotics: Manipulators in manufacturing Robotics: Humanoids http://asimo.honda.com/inside_asimo_movies.asp Robotics: Humanoids http://asimo.honda.com/inside_asimo_movies.asp Robotics: Humanoids http://asimo.honda.com/inside_asimo_movies.asp Robotics: DARPA Grand Challenge 2004 Challenge: 142 miles from Barstow, California to Primm, Nevada $1M prize 15 robots qualified, none finished the race 7.4 miles was furthest distance 2005 Challenge: 132 miles in the Mojave desert $2M prize 24 robots qualified, 5 finished the race Top team: 132 miles in 6 hours, 53 minutes. Robotics: DARPA Grand Challenge Robot connected to a laptop that performs reasoning. The goal of the robot is to play Mastermind autonomously with a human player, with the robot trying to break the secret pattern. The robot must sense the game state, decide its actions, and control the game pieces in real time. MS student project at CS/UCLA using LEGO's Mindstorm Kit Faculty at CS/UCLA doing AI-related research Adnan Darwiche Mike Dyer Petros Faloutsos Rich Korf Judea Pearl Stefano Soatto Demetri Terzopoulos Song Chun Zhu Thank you! ...
View Full Document

This note was uploaded on 05/03/2010 for the course CS 1 taught by Professor Rohr during the Spring '08 term at UCLA.

Ask a homework question - tutors are online