Games as a Test Bed for Developing AI Applications (in Physics) Brains vs Computers Symposium of the “van der Waals” study association December 10, 2013 Jos Uiterwijk Department of Knowledge Engineering Maastricht University 1/36
Dec 16, 2015
Games as a Test Bed for Developing AI Applications (in Physics)
Brains vs Computers
Symposium of the “van der Waals” study association
December 10, 2013
Jos UiterwijkDepartment of Knowledge Engineering
Maastricht University1/36
• Computer chess and computer games
• The role of computer games in Artificial Intelligence
• Brute force?
• The impact of knowledge and heuristics
• New developments
• AI and Physics
• Conclusions
Jos Uiterwijk
Overview
2/36
Jos Uiterwijk
Some history
• Start of computer chess
• The Turk
• It was all fake!
3/36
Jos Uiterwijk
Origin of the AI• Starts around 1950
• Chess as the drosophila melanogaster of AI
• 2 pioneers:– Alan Turing
– Claude Shannon
4/36
1. Rules are simple, but the strategy is complex
2. Domain is fixed, by which programs are easily comparable, both with other programs and with humans. By the nature of a game it is easy to test if a new technique is “better”.
3. Games are typical for human intelligence (Goethe: chess is the touchstone of the intellect). This explains the interest from psychology.
Jos Uiterwijk
Why is chess of interest for AI?
5/36
Jos Uiterwijk
Alan Turing (1912-1954)
• Worked in Bletchley Park during World War II
• Decoding the German Enigma codes: The Bombe
• Was the first who seriously posed the question: Can machines think?
6/36
Jos Uiterwijk
Alan Turing: Turing test
7/36
• Used the computer in his “spare time” for chess programming.
• Was the first who wrote a chess program
Jos Uiterwijk
Alan Turing
8/36
Jos Uiterwijk
Claude Shannon (1916-2001)• Was also concerned with
computer chess
• Built chess endgame machines
• Wrote the “bible” of chess programming: Programming a computer for playing chess (1950)
9/36
Jos Uiterwijk
Claude Shannon• Shannon was
brilliant in many domains, both theoretically and practically: he built among others a juggling robot
• By the way, he also could juggle himself quite good!
10/36
Jos Uiterwijk
State of the Art in computer chess
• Nowaday computers are stronger than human world champions
• Mile stone: Kasparov losing from chess machine Deep Blue in 1997
• Kramnik loses from the “simple” desk top program Deep Fritz in 2006
11/36
Jos Uiterwijk
Adriaan de Groot (1914-2006)
• Professor in psychology
• Studies on “chess thinking”
• PhD thesis (1946) Het Denken van den Schaker, translated (1965) as Thought and Choice in Chess
12/36
• For many board games just used brute computer power was initially used (i.e., as many calculations as possible): dumb but fast. This is called the brute-force approach
• Later the question arose: can the brain still beat the machine by clever use of knowledge? The knowledge-based approach
Jos Uiterwijk
Brute force
13/36
• Facts
• Heuristics(rules of thumb)
Jos Uiterwijk
Knowledge
14/36
• The “mutilated chess board” problems:Can I put domino stones (of 2 x 1 size) in such a way on the board that all squares are covered?
Jos Uiterwijk
Facts (1)
??
The 4x4 problem
15/36
The 8x8 problem
Jos Uiterwijk
Facts (2)
16/36
The 20x20 problem
Jos Uiterwijk
Facts (3)
17/36
• The knight jump puzzle:Can you find a route on the chess board starting at a
given location such that all squares are traveled exactly once?
• Heuristic (rule of thumb):First visit the corners, then the edges, etc., gradually
going to the centre
• Does this heuristic work?
Jos Uiterwijk
Heuristics (1)
18/36
• The knight jump puzzle on the 8x8 board
Jos Uiterwijk
Heuristics (2)
45 32 11 16 43 34 14
10 17 44 33 12 15 42 35
31 46 59 56 61 52 13 2
18 9 62 53 58 55 36 41
47 30 57 60 51 64 3 24
8 19 50 63 54 25 40 37
29 48 21 6 27 38 23 4
20 7 28 49 22 5 26 39
19/36
• However: a heuristic is fallible:
• The 5x5 knight jump puzzle
Jos Uiterwijk
Heuristics (3)
20 11 6 18
5 16 19 12 7
10 21 ? 17 2
15 4 23 8 13
22 9 14 3 24
20/36
• Another knight jump problem
Jos Uiterwijk
Chosing the right representation
What is the shortest route to switch the white and black knights?
start goal
21/36
• Step 1: number the squares:
Jos Uiterwijk
Solution
108 95 6 7
1 2 3 4
22/36
• Step 2: draw the neighbour diagram for knight jumps (which squares are reachable in one jump?)
Jos Uiterwijk
10 8 9 5 6 7
1 2 3 4
310 6 1 8 7 2 9 4 5
from:
we get:
23/36
• Step 3:
Jos Uiterwijk
B W W B W B B W
and the goal situation
Draw the start situation
24/36
• Step 4:
Recognise that this is just a
railcar switching problem!
Jos Uiterwijk25/36
• Step 5: solution now is simple:
Jos Uiterwijk
B W W B B
W W B B B W W
W B B W
B B W W
12 steps
14 steps
6 steps
8 steps=====
40 steps
26/36
• Many other games have been or are target of AI research.
• Many have been solved, which means that the computer has an optimal strategy against any resistance.
• Others are played above human level
• Some are still difficult
Jos Uiterwijk
State of the art for other computer games
27/36
• Until recently, strong humans refused to play against computers
• Reason: computers are too strong• (recently the human world champion nevertheless
played a match; he lost 8-0!)
Jos Uiterwijk
Reversi
28/36
• Standard boards are completely solved• The Checkers program CHINOOK even gained the
official World Champion title
Jos Uiterwijk
Connect-Four, Domineering, Checkers
29/36
• Until recently, strong humans refused to play against computers.
• Reason: humans are too strong!
• This is, since 10 years, rapidly changing though (Monte Carlo simulations work great!)
Jos Uiterwijk
Go
30/36
• Much progress on:– Games with chance (Poker, Backgammon)– Multi-player games (Chinese checkers)– Imperfect-information games (Bridge)– Real-time strategy (RTS) games
Jos Uiterwijk
And many other games ...
31/36
• Just mentioning some physics domain:– Nuclear physics (safety!) – Medical physics (data analysis; data mining)– Robotics (large progress)– Vision (still difficult)– Intelligent design– Many kinds of simulations, from microscopic
small to astronomic large
Jos Uiterwijk
Applications in Physics
32/36
• They all benefit from computer science and AI in particular, such as:– Fast calculations– Machine learning– Pattern recognition / data mining
Jos Uiterwijk33/36
• Chess (and other games) – Drosophila melanogaster of AI– Tool for development of new techniques– Insight into human intelligence
• Computers can play many (board) games at (supra) expert level
• Other games are still a challenge (Go).
Jos Uiterwijk
Conclusions
34/36
• Can and will computers beat the human brain?– Yes, in many complex (but otherwise dumb)
domains– No, in several not so complex, but intelligent,
domains, for a long time to go!– Much game research has to be done!
Jos Uiterwijk35/36
Jos Uiterwijk
The End!
36/36