Solving Tsumego On Computers: M2 Hirokazu Ishii Chikayama & Taura Lab

Download as ppt, pdf, or txt
Download as ppt, pdf, or txt
You are on page 1of 34

Solving

Tsumego on
Computers
M2 Hirokazu Ishii
Chikayama & Taura
Lab

Agenda
1.
2.

3.

Introduction
Related Work
GoTools
Tsumego Solver

4.

Df-pn (Depth-First Proof-Number Search)


Df-pn+ (Depth-First Proof-Number Search+)

Conclusion & Future Work

Agenda
1.
2.

3.

Introduction
Related Work
GoTools
Tsumego Solver

4.

Df-pn (Depth-First Proof-Number Search)


Df-pn+ (Depth-First Proof-Number Search+)

Conclusion & Future Work

Computer Game Players

Have been studied for many years


Easy

to evaluate
Suitable to try out various basic technologies

Were successful in some games


Othello,

Great challenges remain in some games


Shogi,

backgammon, chess

Go

I study on Go, especially tsumego.

Game Rule

Go is a 2-players game.
A move is played on a line intersection.
If one or more stones are completely surrounded
by other stones, they are captured.
The purpose of Go is to conquer a larger part of
the board than the opponent.

Game Rule

Eye
An

eye is an area within a group of stones


which is completely surrounded by stones of
the group.
If the group gets two eyes, they are absolutely
alive.

Definition of Tsumi

Variation of Tsumi
Semeai

problem
Escape and Disconnect problem
Life and Death problem

Goal is to distinguish whether there are two eyes


or not.

Method

AND/OR tree search


OR

NODE

It corresponds to a first players move.


In order to prove tsumi of this node, we must prove
that one node of child nodes has tsumi.

AND

NODE

It corresponds to a second players move.


In order to disprove this node, we must disprove
that all nodes of child nodes are tsumi.

Agenda
1.
2.

3.

Introduction
Related Work
GoTools
Tsumego Solver

4.

Df-pn (Depth-First Proof-Number Search)


Df-pn+ (Depth-First Proof-Number Search+)

Conclusion & Future Work

GoTools
Has been the best tsume-go solver for 15
years.
Uses a depth-first search.
Specializes in completely enclosed
positions.

Heuristics

Static rules
Most

of the heuristic rules are static.

It is at a particular auspicious point.


It completes one or more eyes.

Static

rules rate moves lower or higher.

Dynamic rules
The

moves refuting opponent moves at


subsequent positions also get some credit .

Dynamic Rules

1A

2B

3B
4C

Number
The

sequence
in which the
moves are
made.

6D

5D
:OR NODE
:AND NODE

Letter
The

field where
the stone is
placed.

Agenda
1.
2.

3.

Introduction
Related Work
GoTools
Tsumego Solver

4.

Df-pn (Depth-First Proof-Number Search)


Df-pn+ (Depth-First Proof-Number Search+)

Conclusion & Future Work

Tsumego Solver

Important factors are to


Recognize

the position
Generate candidate moves
Evaluate these moves and select the next move

Evaluation of a tsumego is strictly


determined life or death.
The only method of finding a strict solution
is to search.

Df-pn Search

Depth-First Search
Memory

and time are used efficiently.


It is necessary to set a threshold.
Proof number
Disproof number

Best-First search

Iterative Deepening
Searches

threshold.

are tried iteratively increasing the

Proof Number
Proof number is the minimum number of
descendant nodes which must be proven
in order to prove the node.
We can see proof number of the node as
the minimum resource required for search.

Proof number is effective for search


because we want to search the most
promising node.

Disproof Number
Disproof number is the minimum number
of descendant nodes that must be
disproven in order to disprove the node.
Disproof number is effective for search as
well as the proof number.

Df-pn Search

If n is a leaf node
when

the value is true

when

the value is false

pn(n) 0, dn(n)

pn(n) , dn(n) 0

When n is an uninspected node


The

node might be proven or disproven


immediately when inspected

pn(n) 1, dn(n) 1

Df-pn Search

If n is an internal node
when

n is an OR node

pn(n)
dn(n)
when

min

Cchildlen of n

pn(C )

dn(C )

Cchildren of n

n is an AND node

pn(n)

pn(n )

Cchildren of n

dn(n)

min

Cchildlen of n

dn(C )

Df-pn Search
Is a depth-first search.
Uses two kinds of threshold (proof number
and disproof number)

.Assign

r.th p , r.thd
where r is the root node

Df-pn Search
.At each node n, the search process
continues to search below n until

pn(n) n.th p
or

dn(n) n.thd
(We call it the ending condition)

Df-pn Search
.If n is an OR node
At

each node n, select the child nc with the


minimum proof number and the child n 2 with
the second minimum proof number. Search
below nc with assigning

nc .th p min(n.th p , n2 . pn 1)
nc .thd n.thd nc .dn
C.dn
Cchildren of n

Df-pn Search
.If n is an AND node
At

each node n, select the child nc and the


child n2 . Search below nc with assigning

nc .th p n.th p nc . pn

C. pn

Cchildren of n

nc .thd min(n.thd , n2 .dn 1)

.If the ending condition holds, the search


process returns to the parent node.

Df-pn Search

[ pn, dn]
(th p , thd )

:OR NODE
:AND NODE

(2,-2)

30 nodes

[1,1]
[30,1]

[1,1]
[0,]
[1,3]
[2,2]

(31,-1)
(2,-2)

[1,1]
[2,1]
[0,]

(3,-2)

[1,1]
[,0]

(30,2)
D

[1,1]

(,)

[0,]
[1,1]
E
(31,-1)

[0,]
[1,1]

[1,1] [1,1]
H
[0,]

[,0]

[0,]

[0,]

[,0]

Df-pn+
Intends to distinguish promising moves
more accurately and to search them much
more deeper.
Uses two kinds of additional information
during search.

Df-pn+

Two kinds of information


cost(dis)proof(n,

nchild)

The cost of inspection of nchild starting from n.

h(dis)proof(n)

Heuristic estimate of the cost to reach any proof


solution from position n.

Df-pn+

The formula for calculating proof number


and disproof number are modified from
df-pn.

If node n is an uninspected node

pn(n) h proof (n)


dn(n) hdisproof (n)

Df-pn+

If n is an internal node
pn(n)

dn(n)

min

Cchildlen of n

( pn(C )

(dn(C )

Cchildren of n

cost proof (n, C ) )


cost disproof (n, C ) )

At each node n, search below nc with


assigning
nc .th p min(n.th p , n2 . pn cos t proof (n, n2 ) 1)
cos t proof (n, nc )
nc .thd n.thd nc .dn (nchild .dn cos t disproof (n, nc ))

Bouzys 5/21 Algorithm

Dilation
If

the intersection is not surrounding


opponents stones, then add to the
intersection the number of own interim
territory and stones.

Erosion
Subtract

the number of intersections


with opponents stones or vacant.

Bouzys 5/21 Algorithm


1
2

2
4

1
2

1
2

4 Dilation + 2 Erosion

10

10

5
12

11

12

1
5

10

Bouzys 5/21 Algorithm


In

GNU Go ver. 2.6, it was


extensively used
5

Dilation and 21 Erosion are used for territory.


5 Dilation and 10 Erosion are used for moyo.
4 Dilation and 0 Erosion are used for area.

Higher Bouzy value means higher


expectation to form an eye.
Because two eyes are required to live, the
second maximum Bouzy value may give a
good criterion.

Agenda
1.
2.

3.

Introduction
Related Work
GoTools
Tsumego Solver

4.

Df-pn (Depth-First Proof-Number Search)


Df-pn+ (Depth-First Proof-Number Search+)

Conclusion & Future Work

Conclusion & Future Work

I showed
Definition

of tsumi
Feature of tsumego solver GoTools
Df-pn+
Bouzys 5/21 algorithm and its use in tsumego.

Firstly, I have to finish implementing the program.


I also plan to apply the combinatorial game theory
to the program.

You might also like