Artificial Intelligence AI Topics History and Overview
- 2. AI Topics History and Overview Machine Learning Games and AI The Turing test Computer Vision
- 3. AI Pioneers Alan M. Turing “ Computing Machinery and Intelligence” Marvin Minksy Constructed the first neural net machine Herbert Simon, Allen Newell, J.C. Shaw Developed the first AI computer program
- 4. AI Terms Artificial Intelligence: The capability of a machine to imitate intelligent human behavior Artificial Neural Network: A network of neurons with connections of varying strength Fuzzy Logic: A superset of Boolean logic which includes truth values between true and false Knowledge Base: A collection of knowledge expressed using some formal knowledge representation language AI-complete: Describes a problem which presupposes a solution to the “strong AI problem”
- 5. Famous AI Programs ELIZA (Joseph Weizenbaum) Psychologist Deep Blue (IBM) Chess program Cyc (MCC and Cycorp) Multi-contextual knowledge base and inference engine HAL (Arthur C. Clarke) Space explorer
- 7. What Is Machine Learning? Enabling machines to process data in such a way that it can be to make future decisions ML been studied for many years ML has many applications in a variety of fields
- 8. Methods of Learning Genetic algorithm Inductive logic Computational learning
- 9. Dimensions of Study Representation of experience Most learning is based on experience Storage values Attribute values (length) Binary values (yes/no) Relations (Difficult) Representation of acquired knowledge Generalizations Logical/discrete vs. numeric/continuous
- 10. Dimensions of Study Supervised and unsupervised learning Supervised Feedback given immediately after an action is taken Easy to give examples of correct vs. incorrect behavior Unsupervised Machine learns on its own with no conditioning Inductive learning vs. analytic learning Inductive – take all data, make generalizations Analytic – offer explanations for new data based on previous data, then simplify
- 11. Dimensions of Study Incremental vs. Non-Incremental Learning Incremental Examine results one-by one Less information retained, but faster Non-Incremental Examine all results at once More information retained, but slower
- 12. Tasks For Machines Pattern recognition Grouping/classification Create general descriptions for classes of instances Strategizing Generating heuristics Problem solving
- 13. Problem Solving Take a similar problem with a known solution and try to find the answer (analogies) Simplify the problem and find a solution that can be used to solve the main problem Thresholds Decision trees Macro-operators (AND, OR)
- 14. Issues in Machine Learning Computational complexity Ethics Correctness Would the exact desired learning be constructed? What if there is an error in learning?
- 15. Games AI Min-Max Trees Builds a level of maximizing moves followed by a level of minimizing moves Uses evaluate functions to analyze situation Alpha Beta Trees Like Min-Max Trees Discards paths it knows to be useless
- 16. Chess Algorithms Most use Alpha-Beta trees to make moves Trees helped by additional knowledge Transposition Tables Endgame Database Human Literature Deep Blue First championship caliber chess player
- 17. Other Games Othello – Logistello Deep search algorithm Can solve most endgames Large opening book Checkers – Chinook Extremely deep search depth 8 piece endgame database
- 18. The Turing Test Motivated to identify intelligence in a computer program. Proposed in 1950 by Alan Turing. Original Proposal: Given a person X, a computer Y, and an interrogator C, C isolated from X and Y. C must determine who is the person X is intelligent if it can fool C.
- 19. Problems with the Turing Test Intelligence may be considered as a continuum. The Turing test only identifies one (very strong) type of intelligence, and thus offers no means to measure . Does fooling C really imply intelligence?
- 20. Our Proposal Motivated to allow: a measure of intelligence. more rigid definitions. more flexible admission of programs.
- 21. Our Proposal Define D as the set of all problems. This may be restricted for practical considerations. P(D) is therefore the partially ordered set (under inclusion) of all subsets of problems.
- 22. Our Proposal Let R be the set of all responses P(R) is therefore the partially ordered set of subsets of R. Define the Turing Test T as a function between P(D) and P(R). Those programs which mimic T on some subset X (pre-image) of P(D) are said to pass T restricted to X .
- 23. Our Proposal As P(D) is partially ordered, and by the way D was defined, there are several maximal elements M i in P(D). A program that is said to pass T restricted to an M i is said to be an expert in M i . In specific applications, one may identify an expert program as intelligent.
- 24. Examples Consider the set of Arithmetic Problems If a program can solve these problems, it is said to pass T restricted to Arithmetic Problems . In practice, one would need to restrict this set.
- 25. Examples The set of all Math Problems is a maximal element. If a program can solve these problems, it is said to be an Expert in Math Problems .
- 26. Sources Encyclopedia of Artificial Intelligence 2 nd ed. Ed. Stuart C. Shapiro. John Wiley & Sons, Inc. New York City, NY, 1992. R. Miikkulainen and D. Moriarity. Discovering Complex Othello Strategies Through Evolutionary Neural Networks. University of Texas, USA, 1995. B. Moreland. Basic Search Techniques. http://www.seanet.com/~brucemo/topics/topics.htm. USA, 2001. J. Schaffer. The Games Computers (And People) Play. University of Alberta, Canada, 2000