Chapter 32. Extraordinary Achievements

Transcription

1 Chapter 32. Extraordinary Achievements The Quest for Artificial Intelligence, Nilsson, N. J., Lecture Notes on Artificial Intelligence, Spring 2012 Summarized by Kim, Kwon-Ill and Yoo, Jun Hee Biointelligence Laboratory School of Computer Science and Engineering Seoul National Univertisy

2 32.1 Games Chess Checkers Other Games 32.2 Robot Systems Contents Remote Agent in Deep Space Driverless Auto mobiles 2

3 32.1 Games Overview of Chapter 32 AI applications for playing games Heuristic search & learning 32.2 Robot Systems Automatic control systems Space ship Automobiles 3

4 Chapter 32. Extraordinary Achievements 32.1 Games 4

5 32.1 Games Computers is thought by some to be a some what frivolous diversion from more serious work. Computer game-playing has served as a laboratory for exploring new AI techniques especially in heuristic search and in learning. 5

6 Chess Deep Blue In 1997, the world chess champion Garry Kasparov has been defeated by IBM s Deep Blue (2 win/1 lose/ 3 draw). Deep Blue uses heuristic search. Is this victory can be said AI achievement? IBM s opinion No Figure 32.1: Garry Kasparov playing chess against Deep Blue in game two of a six-game rematch. Chess 6

7 Chess Differences between Kasparov and Deep Blue Deep Blue Kasparov Evaluation per second 200,000,000 3 Amount of Chess knowledge Small Large Calculation ability Huge Small Used skills Search Tremendous sense of feeling and intuition Adaptive Thinking Can t Very quickly Although Deep Blue won, it s not sort of AI. - IBM 7

8 Checkers is Solved Checkers In September 2007 Jonathan Schaeffer published an article about Checkers game. It was announcing that Perfect play by both sides leads to a draw. There are 500,995,484,682,338,672,639 different positions in Checkers. Along the way to the proof, Jonathan s team developed a checkers program named CHINOOK. Figure 32.2: Jonathan Schaeffer. Checkers 8

9 Poker Other Games Heads-Up Texas Hold em (Limit and No Limit). University of Alberta (2008) Bridge Ginsberg s Intelligent Bridgeplayer Go (a.k.a Ba-Duk) One of the hardest game for computers Scrabble Especially suited for computers International Computer Games Association (ICGA) 2011, SNU CSE Biointelligence Lab., 9

10 Chapter 32. Extraordinary Achievements 32.2 Robot Systems 10

11 32.2 Robot Systems Robots are everywhere! For Mars, deep ocean, volcano Agricultural robots, factory robots, surgical robots, and warehouse robots More than 30 robotics companies in Pittsburgh But, it is not intelligent. Remote control by human Improving to autonomous, intelligent action 11

12 Remote Agent in Deep Space 1 Deep Space 1(DS1) Oct 24, 1998, NASA Remote Agent (RA) Robotic system for planning and executing Ex) During the next week take pictures of the following asteroids and thrust 90% of the time By Jet Propulsion Laboratory Programmed in LISP Works RAX: Space-tested version RA Figure 32.4: Artist's rendering of DS1 approaching a comet. Figure 32.6: Illustration of RAX activities. 12

13 Driverless Automobiles Very challenging!! Rapid planning and reaction Wide range of conditions On sunny and stormy days, at night, on city streets, on high-speed motorways, and on and o desert roads Crashes In USA, 28,933 people died & 2,221,000 injured in 2007 But, this numbers mean only 1 person killed per 100 million vehicle miles traveled!! 13

14 Driverless Automobiles DARPA s Grand Challenge in 2004 Auto-drive on-and-off road in the desert ALVINN, RALPH, and Navlab by CMU All failed DARPA s Grand Challenge in team completed the course!! Computer vision technology Figure 32.7: Stanley on Beer Bottle Pass followed by a DARPA chase vehicle. Figure 32.8: Sandstorm on Beer Bottle Pass. 14

15 Driverless Automobiles DARPA s Urban Challenge in mile course in a mock city environment 6 team completed successfully!! Future of Driverless Automobiles By 2030, half of our highway miles will be driven autonomously without human input. by S. Thrun Various societal and legal problems Automated aids to human drivers in a few years Figure 32.10: Tartan Racing team leader William (Red) Whittaker and Boss pose with first place trophy. 15

16 32.1 Games Chess Checkers Other Games 32.2 Robot Systems Summary Remote Agent in Deep Space 1 Driverless Auto mobiles 16

17 Chapter 32. Extraordinary Achievements Appendix 17

18 IBM s opinion Chess Deep Blue uses brute-force methods. Deep Blue could draw on standard moves over 4,000 positions and also be influenced by a 700,000 grand master game database. No formula exists for intuition. 18

19 Chess In broader view of A.I. another view Although Deep Blue relied mainly on brute-force methods (rather than on rule-based reasoning), it use heuristic search (one of foundational techniques of AI). In 2006 World Chess Champion Vladimir Kramnik vs. Deep Fritz Deep Fritz (version 8) won 2 games and 4 draws. The latest version of Deep Fritz is

20 Checkers 20

21 Checkers Figure 32.3: Schematic for the checkers proof. The Proof Axis y: the number of rest pieces. Axis x: the logarithm of the number of positions. Shaded area: less than 10 pieces left (39,271,258,813,439 positions) Optimum play involves using heuristic search to find a line of play guaranteed to get shaded area. 21

22 Checkers vs. Human In 1992, Checkers champion Marion Tinsley beat CHINOOK four wins to two, with thirty-three draws. Rematch has been held in 1994, CHINOOK was declared the Man-Machine World Champion because of Tinsley s resign, citing health reasons. Figure 32.2: Jonathan Schaeffer. Checkers 22

23 Other Games Poker Heads-Up Texas Hold em (Limit and No Limit). University of Alberta (2008) Devoted to the AAAI Computer Poker Competition. Royal straight flush 23

24 Other Games Bridge Goren in a Box (a.k.a Ginsberg s Intelligent Bridgeplayer) Matt Ginsberg Uses Monte Carlo approach. Bridge play table 24

25 Go (a.k.a Ba-Duk) MoGo Titan (2008) Other Games INRIA France and Maastricht University in the Netherlands. Beat a professional Ba-Duk player in a game with the Dutch supercomputer Huygens. This was given a handicap of nine stones. Ba-Duk is probably one of the hardest game for computers. Human vs. Computer Go games: Ba-Duk table 25

26 Scrabble Other Games Especially suited for computers with their abilities to access large dictionaries and conduct massive searches. Scrabble programs now routinely beat expert humans. Brian Sheppard, World-Championship-Caliber Scrabble," Artificial Intelligence, Vol. 134, Nos. 1-2, pp , January Scrabble board 26

27 Other Games International Computer Games Association (ICGA) Information about all kinds of computer game-playing tournaments. Homepage of ICGA 27

28 Remote Agent in Deep Space 1 Subsystems of RA Planner/Scheduler (PS) Mission Manager (MM) Smart Executive (EXEC) Mode ID system Procedure Given mission goal & spacecraft state MM formulates a planning problem for PS PS construct a plan (schedule of actions) for EXEC Planning Experts participate in planning EXEC decompose high-level schedule to commands for Real-Time Execution For failed tasks, EXEC attempt alternatives or Mode ID and Recovery system analyze & repair the problem Figure 32.5: Remote agent architecture. 28

29 Driverless Automobiles Brief history ALV project of the Strategic Computing Program in the mid-1980s by DARPA ALVINN, RALPH, and Navlab by CMU VaMP by E. Dickmanns at Universitat der Bundeswehr in Munich Drive from Munich to Odense, Denmark, and back in 1995 Computer vision and radar Tsukuba Mechanical Engineering Lab in Japan 2getthere in Netherlands ARGO Project in Italy 29

30 Driverless Automobiles DARPA s Grand Challenge in 2004 Auto-drive on-and-off road in the desert For unmanned aircrafts & vehicles for army 142-mile in 10 hours 2000 waypoints, navigating around obstacles, staying on roads, avoiding drop-offs, GPS system $1 million prize All failed Farthest travel : 7.5 mile Some vehicles, good at following way points, are poor at avoiding obstacles, and vice versa. 30

31 Driverless Automobiles DARPA s Grand Challenge in team completed the course!! 1 st : Stanley from Stanford University Used ranging and optical sensors Computer vision technology Figure 32.7: Stanley on Beer Bottle Pass followed by a DARPA chase vehicle. Figure 32.8: Sandstorm on Beer Bottle Pass. 31

32 Driverless Automobiles Technologies in Stanley, the winner of GC2005 Sebastian Thrun, Michael Montemerlo System Six-processor computing platform (Intel) Drive-by-wire control system Sensors 5 laser range-finding units, 1 video camera, GPS system, Gyroscope, Accelerometers Probabilistic terrain analysis (PTA) Distinguish drivable / nondrivable terrain Computer vision Drivable surface identification with surface patches Online speed control Trade off risk and speed Figure 32.9: Sebastian Thrun (left) and Michael Montemerlo (right). 32

33 Driverless Automobiles DARPA s Urban Challenge in 2007 Visiting check points in 6 hours 60 mile course in a mock city environment Merging, passing, parking, negotiating intersections, California driving regulations Traffic: 50 vehicles simultaneously 89 applicants 11 teams for final 6 team completed successfully!! 1 st : Boss from CMU Figure 32.10: Tartan Racing team leader William (Red) Whittaker and Boss pose with first place trophy. 33

34 Driverless Automobiles Technical issues in Urban Challenge Follow rules of the road Detect and track other vehicles at long ranges Find a spot and park in a parking lot Obey intersection precedence rules Follow vehicles at a safe distance React to dynamic conditions such as blocked roads or broken-down vehicles Competitions sponsored by companies Volkswagen in 2007 GM in

35 Driverless Automobiles Future of Driverless Automobiles By 2030, half of our highway miles will be driven autonomously without human input. by S. Thrun Various societal and legal problems Accident liability Human desire to be in control Automated aids to human drivers in a few years All-around collision warning systems Radar-based cruise control Lane-change warning devices Electronic stability control GPS & digital maps 35