124. Chew (3k) vs Fuego (Bot)

User avatar
emeraldemon
Gosei
Posts: 1744
Joined: Sun May 02, 2010 1:33 pm
GD Posts: 0
KGS: greendemon
Tygem: greendemon
DGS: smaragdaemon
OGS: emeraldemon
Has thanked: 697 times
Been thanked: 287 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by emeraldemon »

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Unhidden:

Here's the first few moves of the sequence:

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . 5 . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , 4 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 6 . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


The only other move it seriously considered was this:
Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . 3 . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X . . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
Last edited by emeraldemon on Mon Jun 13, 2011 2:40 pm, edited 1 time in total.
User avatar
emeraldemon
Gosei
Posts: 1744
Joined: Sun May 02, 2010 1:33 pm
GD Posts: 0
KGS: greendemon
Tygem: greendemon
DGS: smaragdaemon
OGS: emeraldemon
Has thanked: 697 times
Been thanked: 287 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by emeraldemon »

perceval wrote:Out of curiosity, fuego proabably has some routines to select the next move at least.
Do you know how far they go and how many "rule based candidate" it is considering ?
meaning, did it just select n candidate and evaluate them directly or does its rule engine also construct a few W answers move before turning to MC evaluation of each ?



There are some 1st line moves it removes by policy, but everything else is selected directly by UCT search.
User avatar
daniel_the_smith
Gosei
Posts: 2116
Joined: Wed Apr 21, 2010 8:51 am
Rank: 2d AGA
GD Posts: 1193
KGS: lavalamp
Tygem: imapenguin
IGS: lavalamp
OGS: daniel_the_smith
Location: Silicon Valley
Has thanked: 152 times
Been thanked: 330 times
Contact:

Re: 124. Chew (3k) vs Fuego (Bot)

Post by daniel_the_smith »

Hm, I'd almost always play this. Is fuego's suggested move better? Discuss! :)

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . 4 . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com
User avatar
Chew Terr
Gosei
Posts: 2060
Joined: Mon Apr 19, 2010 12:45 pm
Rank: KGS 3k
GD Posts: 264
KGS: Chew
Location: Texas
Has thanked: 546 times
Been thanked: 172 times
Contact:

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Chew Terr »

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . 4 . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O a 5 8 O . . . . . X . . . |
$$ | . . . . . O 9 6 7 . . . . 4 . . . . . |
$$ | . . . . . . b 0 . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

This works okay because 'a' and 'b' are miai. Like I said, this approach is great for me. After this, I'm tempted to play on top. I could also take a local point like F5, thought that ends in gote after F6 G5.
Someday I want to be strong enough to earn KGS[-].
ratata
Beginner
Posts: 14
Joined: Wed Apr 21, 2010 11:28 pm
Has thanked: 5 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by ratata »

daniel_the_smith wrote:
Hm, I'd almost always play this. Is fuego's suggested move better? Discuss! :)

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . 4 . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


If black is not able to strongly attack white I prefer Fuego's move, since it creates a double-wing formation.
User avatar
daniel_the_smith
Gosei
Posts: 2116
Joined: Wed Apr 21, 2010 8:51 am
Rank: 2d AGA
GD Posts: 1193
KGS: lavalamp
Tygem: imapenguin
IGS: lavalamp
OGS: daniel_the_smith
Location: Silicon Valley
Has thanked: 152 times
Been thanked: 330 times
Contact:

Re: 124. Chew (3k) vs Fuego (Bot)

Post by daniel_the_smith »

ratata wrote:
If black is not able to strongly attack white I prefer Fuego's move, since it creates a double-wing formation.


Hm? I don't see a double wing in Emeraldemon's diagram?
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com
lightvector
Lives in sente
Posts: 759
Joined: Sat Jun 19, 2010 10:11 pm
Rank: maybe 2d
GD Posts: 0
Has thanked: 114 times
Been thanked: 916 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by lightvector »

Chew Terr wrote:
Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X 3 . . . . . . . . . . . . . |
$$ | . . X , X O a 5 8 O . . . . . X . . . |
$$ | . . . . . O 9 6 7 . . . . 4 . . . . . |
$$ | . . . . . . b 0 . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

This works okay because 'a' and 'b' are miai.


If it goes like this, black can actually eat some stones here.

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O 1 X O O . . . . . X . . . |
$$ | . 9 8 6 4 O X O X . . . . O . . . . . |
$$ | . . . 7 5 3 2 O a . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Also black threatens "a". If black later gets to play here first, is there any way for white to deviate?
User avatar
emeraldemon
Gosei
Posts: 1744
Joined: Sun May 02, 2010 1:33 pm
GD Posts: 0
KGS: greendemon
Tygem: greendemon
DGS: smaragdaemon
OGS: emeraldemon
Has thanked: 697 times
Been thanked: 287 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by emeraldemon »

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . 5 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


I'm still experimenting some with how I can get data from Fuego about its decisions; bear with me while I try some different things and please ask if there's specific info you'd like me to try to find.

First the sequence:

Click Here To Show Diagram Code
[go]$$Bcm15
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . 2 . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . 6 . . . |
$$ | . . . . . . . . . . . . . . 7 . 4 . . |
$$ | . . . X . . . . . , . . . . . 5 . . . |
$$ | . . . . . . . . . . . . . . . . 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . 3 . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . 8 . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


B15 is an interesting choice, leaving enough space around R6 for white to invade later, but threatening to develop on a large scale. B19 and B21 show that Fuego is happy to play for influence. Fuego thinks black has a 49.8% chance to win this game.

I can look at how many times a particular move was visited in the tree, which is a good indication of how much time it spent thinking about that move. Here are all the moves Fuego visited more than 100k times:

Click Here To Show Diagram Code
[go]$$Bcm11 Black's possibilities
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . b . c . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . e . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . a . . |
$$ | . . . . . . . . . . . . . . . . d . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Number of visits:
a: 3,316,615
b: 249,203
c: 144,403
d: 111,646
e: 108,358

A few people (including Chew Terr) commented about H4, but Fuego only visited that move 23 times, so it basically didn't consider it. The only inside attack it thought much about was J3, with 8066 visits.

We can do something similar to see what responses it thinks Chew might play:

Click Here To Show Diagram Code
[go]$$Bcm11 White's possibilities
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . a . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . b . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . 5 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . d . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . c . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Of these, only ''c'' seems unreasonable (looks over-concentrated for this stage of the game).

One note about the choice of R12 (e in diag 2, b in diag 3): Fuego has a strategy called RAVE, which basically says "if a move is good anywhere in the tree, try playing it higher up in the tree". This is a useful heuristic, because it automatically builds in ideas like "my opponent's good move is my good move" and the 1-2-3 rule (if 3 is good, just play that). In this case, I noticed it spent time looking at this sequence:

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 7 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . 5 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . 6 . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


This is a very natural move in this position, and I wonder if the RAVE policy then encouraged both white & black to try R12 earlier in the tree.
Wildclaw
Dies in gote
Posts: 29
Joined: Fri Apr 23, 2010 4:54 pm
Rank: KGS 2 kyu
GD Posts: 543
Has thanked: 3 times
Been thanked: 4 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Wildclaw »

hyperpape wrote: could you elaborate on this point? I had thought that one of the big advantages of Monte Carlo methods is that they make good use of increased computational power, in a way that older programs didn't.


Not really. Min-max can make just as much use of increased computational power. The big advantage of MC is that it is depth first unlike min-max that is width first. That removes the need for a well functioning evaluation algorithm.

A common misconception is that MC playouts are mostly random. It is no more true than saying that human reading is mostly random because we don't go through each path in the exact same order each time. The important part of reading isn't which path you choose first, but which paths you focus most of your concentration on. And the same holds true for MC.
lightvector
Lives in sente
Posts: 759
Joined: Sat Jun 19, 2010 10:11 pm
Rank: maybe 2d
GD Posts: 0
Has thanked: 114 times
Been thanked: 916 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by lightvector »

As a computer science major focusing in AI (and as someone who's followed the computer-go mailing list and the major papers on computer go for a few years), I'd like to add a few comments.

Prior to the development of MCTS (monte-carlo tree search), programs actually didn't make good use of increased computational power. Indeed, for many old-generation programs, such as Gnugo, almost arbitrarily large increases in computational power add hardly any strength improvement at all! This was because old programs lacked a good way to evaluate positions. It didn't matter how deep they could read or how many of the tactics they could figure out, if they couldn't judge the resulting positions.

The "surprise" came when people found that randomly playing it out was actually a pretty good evaluator in Go. Unlike games such as Chess, where randomly shuffling all the pieces will utterly destroy any meaning in the position, in Go, the stones don't move, so that even as you randomly play more moves, the same shapes tend to stick around, and continue to have their effects, whether good or bad. If your group was thick and connected before, it's probably still thick and connected afterwards. If you had good eyeshape before, you probably will still have good eyeshape with a few more stones. If you controlled an area very strongly before, you'll probably still control it very strongly after some random moves.

Basically, the idea is to judge a position by playing it out, while maintaining the "status quo" as much as possible. To that end, a modern go program's playouts are actually very very far from pure random. If the one side during the playout threatens to break a connection, the other side won't play randomly, but will instead connect. If a one side wedges in a one-point jump, then the other side will atari and then connect. If one side plays atari on a group, the other side will connect, or better yet, capture an adjacent group, if possible. If one side plays in the middle of a straight-four eyespace, the other side will respond by playing the other point, to divide it into 2 eyes.

More or less, the playouts will finish the game according to basic instinct. It's not too much of a stretch to imagine why finishing the game according to basic instinct would be a not-terrible way to estimate the value of a position, especially if you do it a lot of times with some random choices of variations and average out the results. Of course, there will be positions that playouts can't handle alone. That's why it's MCTS (monte-carlo tree search), not just MC. You still do the tactical reading nonrandomly in a min-maxing fashion, it's just that you use the above playout method to evaluate the *results*.
Mike Novack
Lives in sente
Posts: 1045
Joined: Mon Aug 09, 2010 9:36 am
GD Posts: 0
Been thanked: 182 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Mike Novack »

Just want to add a bit to what "lightvector" said (I'm from a different area of the trade, but enough computer science to be able to follow)

The devil is in the details.

The problem is that there isn't enough time to do a (real) min-max but that is what these programs are trying to emulate. How can MCTS be "tuned" to converge as close as possible to the min-max solution? The programs are different in how they implement "basic instinct", how they order consideration of moves into the tree, how they prune the tree, etc.

That's why these programs are in fact different. Different in strength, in playing style, etc. and why we should be careful making general statements about them based on our observation of only one program.
User avatar
Chew Terr
Gosei
Posts: 2060
Joined: Mon Apr 19, 2010 12:45 pm
Rank: KGS 3k
GD Posts: 264
KGS: Chew
Location: Texas
Has thanked: 546 times
Been thanked: 172 times
Contact:

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Chew Terr »

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . 6 . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]



Okay, I've said for long enough that I wanted the top move, so here it is. The other move that jumped at me instictively was a knight's move to R14. The biggest problem that I have with that is, the left side is a little small now that my opponent just played low. Because of that, the approach at R14 is less big for black, and any development on the right will be at least a bit small/slow. I need to figure out what, if anything, to do about the left side or bottom left, but at the moment, I am content to just grow my own shape. Something around E13 is looking interesting, as it's at the boundary of frameworks.
Someday I want to be strong enough to earn KGS[-].
Bill Spight
Honinbo
Posts: 10905
Joined: Wed Apr 21, 2010 1:24 pm
Has thanked: 3651 times
Been thanked: 3373 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Bill Spight »

I am going to guess Fuego's next move, based upon my impression of modern Monte Carlo players as influence players in the opening when they go out of book.

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . O . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . 7 . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . . . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


I guess I don't need to hide this. ;)
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.
User avatar
Chew Terr
Gosei
Posts: 2060
Joined: Mon Apr 19, 2010 12:45 pm
Rank: KGS 3k
GD Posts: 264
KGS: Chew
Location: Texas
Has thanked: 546 times
Been thanked: 172 times
Contact:

Re: 124. Chew (3k) vs Fuego (Bot)

Post by Chew Terr »

Yowza, that would definitely shape the game, one way or the other.Be interesting to see, if it happens!
Someday I want to be strong enough to earn KGS[-].
User avatar
emeraldemon
Gosei
Posts: 1744
Joined: Sun May 02, 2010 1:33 pm
GD Posts: 0
KGS: greendemon
Tygem: greendemon
DGS: smaragdaemon
OGS: emeraldemon
Has thanked: 697 times
Been thanked: 287 times

Re: 124. Chew (3k) vs Fuego (Bot)

Post by emeraldemon »

Click Here To Show Diagram Code
[go]$$Bcm11
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . W . . . . . . . . . |
$$ | . . . O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . X . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . 7 . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


Not quite, Bill :)

But maybe soon! Here's Fuego's sequence:

Click Here To Show Diagram Code
[go]$$Bcm17
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 8 . . . . . . O . . . . . . . . . |
$$ | . . 7 O . O . . . , . . . . . O . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . X . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . 6 . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . 4 . . . . . . . . . . |
$$ | . . . X . . . . . , 3 . . . . , . . . |
$$ | . . . . . . . . 2 . . . . . . . X . . |
$$ | . . . . . . . . . 5 . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . X X . 1 . . . . . . . . . . . |
$$ | . . X , X O . . . O . . . . . X . . . |
$$ | . . . . . O . . . . . . . O . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]


You can see Fuego is starting to feel like the center is very important. I doubt Chew will play around tengen like that, so maybe Fuego will take it :)

Fuego predicted Chew to take K17, and the program thought at that time that it should respond with H5. I think it's interesting that it chose H5 again, since it doesn't "remember" the decision from last move (the search tree isn't saved). Seems to imply that the search is mostly consistent.

Also, it's win % has ticked up from 49.8% last move to 50.1% this move, the first time so far it has felt ahead. Who do you think is winning?
Post Reply