AI-Class

1. AI-Class.com Dan Vasicek 2011-10-22

2. Overview of the course (so far) • Problem Solving • Probability in AI • Probabilistic Inference • Machine Learning • Unsupervised Learning • Representation with Logic • Planning • Planning with Uncertainty • Reinforcement Learning • Hidden Markovian Models • Markov Decision Processes • Midterm Exam

3. General information • Huge Student base, ~135,000 students • Publicized in WSJ, and other non-technical sites • Random people that you meet on the street know about the class • This is big enough to have a significant social impact

4. Course Textbook • Artificial Intelligence: A Modern Approach 3rd Ed (1152 pages) • By Peter Norvig and Stuart Russell • Paperback version available from Amazon for $114 • PDF version is available for download for free (45 Megabytes, 7 minutes, & searches for the book name will find the PDF) • CD containing the 45 Megabyte PDF available from Daniel Vasicek

5. Course Mechanics • http://www.ai-class.com– 20 approx 1 hour long lectures presented as 30 or so 2 minute video clips. • Lectures are also available on youtube: http://www.youtube.com/watch?NR=1&v=k-5e935u9hE • During the lectures non-credit quizzes are given • Credit given for homework once per week • Midterm, final exam, and homework count toward your grade • Artistic Intelligence Course at http://www.other-ai.org

6. Planning with Uncertainty Markov Decision Process • http://www.rand.org/pubs/reprints/RP1058.html Carl Dahlman’s Theorem guarantees convergence • Select the policy that maximizes the value of each position for each iteration • Assume we know V(S’) for each location. Then compute a new value maximizing V(S) = Sum(P(S’|S, policy)V(S’) + R(S)

7. Partially Observable Markov Decision Process • Adds information gathering to the MD process

8. Reinforcement Learning • Learn information about the environment • Supervised learning (previous lecture) • Unsupervised learning (previous lecture) • Find an optimal policy when the environment is not known

9. Statistical Independence • Bayes Theorem for events a and b: • P(a and b) = P(a|b)P(b) = P(b|a)P(a) • This gives us a way to evaluate two of the five probabilities when we know three of • P(a|b), P(b), P(b|a), P(a) • Statistically independent means: • P(a and b) = P(a)P(b) = P(b)P(a)

10. Conditional Independence • Conditionally independent means: • P(a|(b and c)) = P(a |c) or • P((a and b)|c) = P(a |c)P(b|c) • Conditional independence does not imply statistical independence • Statistical independence does not imply conditional independence

11. Binary Variables • Battery Age converted from a continuous variable to binary (too old versus young)

12. Bayes Networks • Laplace Smoothing (Not Laplacian !!) • Laplacian smoothing is averaging of adjacent values • Laplace smoothing avoids overfitting by adding weight to unobserved (but possible) cases • Spam Detection • Vocabulary size for Spam versus Ham?

13. K-Means • There seems to be some confusion between K mean values vs K-nearest neighbors

14. K- Nearest Neighbors • Classify unknown points by taking the majority vote of the K nearest neighbor training examples

15. Contact me if you wish • Daniel Vasicek • Cell Phone 505 321 2028 • DanielVasicek@Comcast.net • 2120 Metzgar Road SW 87105 • Land Line 505 873 0575