Wins against strong humans in the game of Go with random initial positions
@inproceedings{helmstetter:inria-00625815,
hal_id = {inria-00625815},
url = {http://hal.inria.fr/inria-00625815},
title = {{Random positions in Go}},
author = {Helmstetter, Benard and Lee, Chang-Shing and Teytaud, Fabien and Teytaud, Olivier and Mei-Hui, Wang and Yen, Shi-Jim},
abstract = {{It is known that in chess, random positions are harder to memorize for humans. We here reproduce these experiments in the Asian game of Go, in which computers are much weaker than humans. We survey families of positions, discussing the relative strength of humans and computers, and then experiment random positions. The result is that computers are at the best amateur level for random positions. We also provide a protocol for generating interesting random positions (avoiding unfair situations).}},
language = {Anglais},
affiliation = {Laboratoire d' Informatique Avanc{\'e}e de Saint- Denis - LIASD , Department of Computer Science and Information Engineering - CSIE , Laboratoire de Recherche en Informatique - LRI , TAO - INRIA Saclay - Ile de France , National Dong Hwa University - NDHU},
booktitle = {{Computational Intelligence and Games}},
address = {Seoul, Cor{\'e}e, R{\'e}publique Populaire D{\'e}mocratique De},
audience = {internationale },
year = {2011},
month = Sep,
pdf = {http://hal.inria.fr/inria-00625815/PDF/randomgo.pdf},
}
1. Random Go
Random Go
1. Many games become boring
2. Randomization brings some fun
3. But computers are terrible in random games
TAO, Inria-Saclay IDF, Cnrs 8623, Lri, Univ. Paris-Sud,
OASE Lab,
Korea,
Summer 2011
1
2. Many games become boring when
they are too much studied
This part for some preliminaries:
1. Go starts from an empty board (usually)
2. Computers are still far from pros in Go
3. Openings are complicated, known, boring (I know
many of you don't agree).
2
3. 1. Go starts from an empty
board
An interesting book on the philosophy of Go:
K. Chairasmisak, “Teachings by an Asian Leader”
(English title by me; exists in French & in Chinese at
Least), 2005. (Thailand CEO of 7-Eleven)
Go has both tactical elements and a strategy
Starting from the empty board is very special
==> link between both ?
==> something specifically human in games
starting from the empty board ?
3
4. 1. Go starts from an empty
board (usually)
Tibetan Go Sunjang Baduk Go Batoo: 3 stones / player
(anywhere)
(maybe value of the
game close to 0.5 ?
whereas 0 or 1 for Go)
===> 3 exceptions; yet, nearly empty
===> and many people study
Tsumegos (not empty board!)
4
5. 2. Computers still far from pros
Go can be played on various board sizes.
9x9 13x13 19x19
5
6. 2. Computers still far from pros
Go can be played on various board sizes.
In 9x9, computers and humans are close.
In 13x13,
computers are still weaker than humans.
H1.5.
In 19x19 Go,
computers are still very weak compared
to pros. H6 is a minimum.
6
7. 3. Openings are boring
Entire books on openings.
Fuseki / Joseki:
Fuseki: full board opening
Joseki: local opening (does not say in which part
of the board you should play)
Less restrictive than in Chess.
Yet, you might have to study a lot if you want to
reach the top level.
7
8. Random Go
Random Go
1. Many games become boring
2. Randomization brings some fun
3. But computers are terrible in random games
8
9. The case of chess
Plenty of complicated openings are known
- We can become strong by rote learning ?
- Is it so great as a pedagogical tool ?
It makes the game a bit boring
Too easy for white to force a draw
Fischer proposed randomization: Fischer Random Chess
(also known as Chess 960)
Principle:
White pieces randomly drawn (within some constraints)
Black pieces in symmetry
9
11. Chess 960: conclusion ?
Strength in Chess960 correlated (but no equal to) strength in Chess
No more boring opening phase
Not many human/computer comparisons
==> we'll do it in Go
Algorithm carefully designed by Fischer for making the game
somehow equilibrated
==> leads to a restricted number of initial situations (960)
==> we'll do this automatically (→ much more initial positions)
11
12. Random Go
Random Go
1. Many games become boring
2. Randomization brings some fun
3. But computers are terrible in random games
12
13. Computers in Go + Random Go
(IEEE SSCI 2011)
The computer The human
Dell Poweredge R900 One big brain.
16 cores, 2.96 GHz, 64bits Ranked fourth in World Amateur
Championship 2004.
Former French champion.
Knows computers.
13
14. Computers vs Humans in
standard Go
Some nice successes in computer-Go
Wins against pros in 9x9
Wins with handicap 6 against pros in 19x19
Yet, the human wins easily with handicap 6.
==> computer crushed
14
15. Computers in random-Go
Randomly put 180
stones on the
board
Can be unfair
So:
Randomly put
180 stones
Check if
equilibrated by
playing multiple
games
If not, restart.
Human chooses
his color.
15
16. Computers vs Humans in
random Go
<= 160 random stones: human wins everything.
240 random stones:
Human chooses black
Computer wins as white
Human wins as white ==> maybe unfair ?
180 random stones:
First position: computer wins both as black and white
Second position: computer looses one game.
Remarks:
Many games the same day ==> human tired
Nonetheless, such results16impossible in standard Go
18. The game of Go is a part of AI.
Consider ridiculous in front of children.
Computers are
a situation with plenty of equivalent
strategies.
Easy situation.
Termed “semeai”.
Requires a little bit
of abstraction.
Random Go Korea 18
19. The game of Go is a part of AI.
Consider ridiculous in front of children.
Computers are
a situation with plenty of equivalent
strategies.
Easy situation.
Termed “semeai”.
Requires a little bit
of abstraction.
All orders for filling
these locations are
equivalent.
All orders for filling
these locations are
equivalent.
Random Go Korea 19
20. The computer stupidly analyzes all
symmetries of all strategies.
800 cores, 4.7 GHz,
top level program.
Plays a stupid move.
Random Go Korea 20
21. The game of Go is a part of AI.
Computers are ridiculous in front of children.
8 years old;
little training;
finds the good move
Random Go Korea 21
22. On which situations are
computers weaker than humans ?
Are there less situations
)
with plenty of invariances
in random-Go ?
Less sophisticated interactions
between various areas of the
board ?
22
23. Conclusions
Computers much stronger from random initial boards.
More fun for humans (see tests in European Go Congress 2011)
New results (not in the paper): tests against “not so strong” humans.
Pedagogy:
- It is often said that Go, Chess, … are good pedagogical tools (true ?).
- But rote learning is important in such games
- Random games a good solution ? (questionable... less human-specific ?)
Computational point of view:
We could not generated positions with many many random stones
==> rejection rate almost 100%
==> How to improve this ?
==> In terms of Go, maybe a good Ishi-No-Shita-generator.
Beyond games:
Generate test cases (for other pbs) with similar approaches ?
Link with testing of some abilities (who plays randomGo well ?)
23