Oski bot

Hi all,

I have written a new Oski-playing bot, "bear_bot". It currently proposes unrated games only.

It will register for rated games or monthly cups if there is enough approval for that.

Please give it a try!

Antony

By the way, it seems that I cannot edit my profile anymore. Do you have the same bug?

There seem to be rules about what you can change to too, in order to avoid spoofing, I think. Good to see another game covered by a bot - which ones are we missing now?

My master just found a nasty bug, I should play better now.

FatPhil: Let's make a list of bots playing here, (and bots widely available).

Go -> fuego

Chess -> (many)

SHOGI -> (many)

Reversi -> (many)

Reversi on board 10x10 -> (many)

Gomoku -> ?

Dots and Boxes -> ?

DVONN -> RoRoRo

Four in row -> ?

Twixt PP -> ?

Hex -> (a few)

Amazons -> (a few)

Breakthrough -> ?

Lines of Action -> MIA (in classic)

Connect6 -> ?

Havannah -> (many)

WYPS ->

OSKI -> bear_bot

Golem word game -> RoRoRo, probably other

StreetSoccer -> nano_ai ?

EinStein würfelt nicht! -> RoRoRo (not backwards)

???

[previous message was a bug and the one before should be read as posted by bear_bot, posted from my account]

So only one guy was interested so far? I feel pretty bad about this :-)

Nice work antony, maybe putting a link will help to attract people a bit lost in LG tree :

here.

I feel better now :-)

Carroll: I will add support for French soon (basically it's just downloading the ODS and writing some extra webpage-parsing code...)

If people want to play in other languages, please tell me what the word list in that language is.

... yet another second-player loss for him. Well done michael!

I really have to implement add a decent lookahead now...

French support now in! Also the code makes it easy for me now to add any other language, as long as you provide the word list to me :-)

Oh merci!

Mais sans "recherche d'arbre" il est déjà trop fort pour moi (au moins en français).

I've done some basic tests with 1-ply lookahead. With either or both players having it, P2 seems still advantaged by approx. 2 pts. Surprising, isn't it?

@antony: There's also Valkyria9 bot playing Go 9x9.

Antony, have you seen the discussions in the thread http://www.littlegolem.net/jsp/forum/topic2.jsp?forum=25&topic=190, where Ed Colins and Richard both made simulations that agreed with this phenomenon without any clear explanation?

Yes, but I use the whole dictionary and can have lookahead (well, actually, I'm not doing a *real* lookahead, but comparing the ~10 words that usually tie for highest scoring in any move). And 10-letter words do occur every so often...

So to be more precise, with 1000 games simulated:

no matter whether none, one or both players use 1-ply (restricted) lookahead, I get ~80% wins for P2, 10% draws and 10% wins for P1; and the mean margin of victory is 2.1 or 2.2 for P2.

I intend to add alpha-beta pruning to my code to do 2-ply (restricted) lookahead in a reasonable amount of time.

Still didn't have the time to add 2-ply lookahead, but some speed improvements are under the way.

I would like to hold a vote about participation of my bot in the MCs and the Championship: who would agree, who would disagree? I intend to enter it with a time limit of a few seconds at most per move. (What are the standards for other bots in other games?)

RoRoRo only takes a few seconds over each move max. 4ir and D&B he's about half a second (and for D&B, that's a perl script, it's not even C). Due to some sloppy coding, he will occasionally go insane and spend up to about 10-15s on a move at Dvonn, even though he's trying to average only a few seconds. (He estimates how effective alpha-beta pruning will be, but sometimes it doesn't deliver the speedup he predicts.)

I'm (in the process of) getting thrashed by michael again... I really have to add lookahead now :P

So I've added a lookahead function. I still assume that it is never good to "sacrifice" a point by playing a non-highest-scoring move (it would be necessary only if it created a TWO-point reduction in the value of the position for the opponent...)

I had it solve completely 1000 games starting from the 8th move, and it used ~4 moves lookahead before the 8th, except for the very first few moves (where the branching factor is the biggest), as P1; P2 followed a greedy strategy of choosing a highest scoring move at random.

Even then, P1 scores only 31.4% of victories, and 24.5% of draws (mean margin of victory is 0.2pt)! That is much better, but means that the game is even more unbalanced than I thought.

"I still assume that it is never good to "sacrifice" a point by playing a non-highest-scoring move (it would be necessary only if it created a TWO-point reduction in the value of the position for the opponent...)"

In our last game I made 'peason' in move 15 instead of 'resales'. 'Peason' gave u only a 6 max response in move 16 while 'resales' would've given u an 8 letter response with 'unreason'. It didn't really matter that much in our game cause I would've still won by 1 point instead of 2. But you can see it would matter if the margin was only 1 point at move 15.

-michael

Nice counter-example. So I have to be more careful indeed :-)

bear_bot is back :-) still in testing though...

bear_bot's strength is currently set to 2-ply lookahead, after some players complained that the initial setting (5-ply) was too strong. Now that the first division championship has been decided (and bear_bot hasn't won :-(), would the players in that division like to play against a stronger bot? Please let me know here.

It's fine for me since you are not dominating the tournament now. But I'm not sure about the others;)

It is ok with me too, I like to play with stronger players, increase the initial setting and let us see :)

Shouldn't we have separate tournaments for those in which bots pariticipate?