CS 331: Artificial Intelligence Systems that Systems that - PDF document

What is AI? (4 categories of defns) Human performance Rationality CS 331: Artificial Intelligence Systems that Systems that Introduction think like Thought process think rationally humans Systems that act Systems that act Behavior like humans rationally 1 2 Problems with the Turing Test Acting like humans (Turing Test) • Not reproducible • Can’t be analyzed mathematically • Tends to focus on human-like errors, linguistic AI Computer tricks, etc. • Does not produce useful computer programs Can a human interrogator, after posing some written questions, tell if the responses come from a human being or a computer? AI researchers believe it’s more important to study the underlying principles of intelligence than duplicating how Requirements for computer: natural language processing, knowledge representation, automated reasoning, machine learning, vision and humans act robotics (the last two are for the “total Turing Test”) 3 4 Thinking Humanly (Cognitive Modeling) Thinking rationally (Laws of Thought) • Models of the internal workings of the Facts and rules Theorem Prover in formal logic human mind • Validation: • Rational = conclusions are provable from inputs and prior knowledge – Compare models with actual behavior of human • Ensure all actions by a computer are justifiable (i.e. subjects (cognitive science) “rational”) – Compare models with neurological activity in the brain (cognitive neuroscience) • AI is now distinct from both cognitive Problems: • Hard to represent informal knowledge formally, especially science and cognitive neuroscience when not 100% certain • Computationally expensive 5 6 1

Rational Agents Acting Rationally (Rational Agents) • “Agent”: something that acts very few resources lots of resources • “Rational” means more than just logically justified. It also means “doing the right limited, no thought Careful, deliberate approximate “reflexes” reasoning thing” reasoning • “Rational agent”: an agent that acts to • Adjust amount of reasoning according to achieve the best outcome given its resources available resources and importance of the result • This is one thing that makes AI hard 7 8 AI Timeline AI Today 1943-1956 The gestation of AI • Deep Blue: first computer program to defeat 1956 The birth of AI the world champion in chess (1996) 1952-1969 Early enthusiasm, great expectations 1966-1973 A dose of reality • AlphaGo: master-level performance at Go 1969-1979 Knowledge-based systems (2016) 1980-present AI becomes a successful industry • NavLab: minivan drove itself across the US 1986-present The return of neural networks 1987-present AI adopts the scientific method on its own 98% of the time (1995) 1995-present The emergence of intelligent agents • Google’s self -driving cars 2001 Big Data • Proverb: crossword puzzle solver (1998) 9 10 Other AI applications in the real Surprises in AI Research world • Credit card fraud detection • Tasks difficult for humans have turned out to be “easy” • Medical diagnosis programs – Chess • Computer-assisted surgery – Checkers, Othello, Backgammon • Search engines – Logistics planning • Personalized news sites – Airline scheduling • Collaborative filtering – Fraud detection • Spam filtering – Sorting mail • Disease outbreak detection – Proving theorems – Crossword puzzles • Opponents in video games 11 12 2

Surprises in AI Research AI Courses at OSU • Tasks easy for humans have turned out to be hard. 1. CS331: Introduction to AI (Spring quarter) – Speech recognition • Search – Face recognition • Games – Composing music/art • Knowledge Representation – Autonomous navigation • Bayesian Networks – Motor activities (walking) 2. CS434: Machine Learning and Data Mining – Language understanding (Spring quarter) – Common sense reasoning (example: how many legs • Supervised Learning does a fish have?) • Unsupervised Learning • Reinforcement Learning 13 14 We will discuss: 1. Search 1. Search Uninformed search Informed search 7 2 4 8-puzzle: Beginning with the start state, slide tiles Local search 5 6 horizontally or vertically until you get to the goal state. 8 3 1 7 4 7 2 4 7 2 4 7 2 4 5 2 6 5 6 5 3 6 5 6 8 3 1 8 3 1 8 3 1 8 1 7 4 7 4 7 2 7 2 4 7 2 4 7 2 4 2 4 7 2 4 7 5 6 8 5 6 5 2 6 5 2 6 5 6 4 5 6 1 5 3 6 5 3 6 8 3 1 8 3 1 8 3 1 8 3 8 1 8 1 8 3 1 3 1 15 2. Games (Fully observable) 3. Knowledge Representation • How do you Knowledge Base From this knowledge base, create a program Everyone from Wisconsin is a can we derive the Packer fan to play tic-tac- All Packer fans like cheese following? toe intelligently? Everyone from Wisconsin is • Your professor is a evil Packer fan • What about Your professor is from Wisconsin • You will have a difficult chess? Evil professors have difficult midterm midterms • Your professor does not like cheese 17 18 3

4. Bayesian Networks Example: Learning to classify emails as spam or not spam P(Spam) = 0.88 P(Spam) = 0.28 Private And Confidential Professor Hutchinson, Dear Friend, I tried to hand in homework 1 electronically It is with heart of hope that I write to seek your but the handin script was broken. I’ve help in the context below. I am Mrs. Jumai attached my homework in this email… Asfatu Abacha, the second wife of the former Nigeria head of state who died on the 8th of June, 1998. Having gotten your address through the internet, I have no doubt on your goodwill to assist us in receiving into your custody (For Safety) the sum of Forty-Eight Million, Five Hundred Thousand United States Dollars (US$48.5M) willed and deposited in my favour by my Late husband… 4