Author |
Topic: 2010 Postal Mixer (Read 12987 times) |
|
rbarreira
Forum Guru
Arimaa player #1621
Gender: 
Posts: 605
|
 |
Re: 2010 Postal Mixer
« Reply #60 on: Jun 24th, 2010, 12:38am » |
 Quote  Modify
|
Fritzlein would you say that if a bot could play like this in a game with 2 minutes per move, it would have a chance to win the Arimaa challenge? Given that humans are probably thinking for longer than 2 minutes about their moves.
Anyway I don't think you have to worry much about 2011. If clueless and Opfor are thinking 2-4 hours, that's 6 to 7 doublings of CPU power, which is a lot to overcome by improving the software.
But my guess is that with improvements to the bots, and the ever-accelerating hardware, by 2020 bots will probably be playing lot better in the challenge than clueless and Opfor are playing right now in the postal tournament.
Of course as you said, there's always the even more uncertain question of human improvement...
|
| « Last Edit: Jun 24th, 2010, 12:39am by rbarreira » |
 IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #61 on: Jun 24th, 2010, 12:49am » |
Quote Modify
|
on Jun 24th, 2010, 12:38am, rbarreira wrote:| Fritzlein would you say that if a bot could play like this in a game with 2 minutes per move, it would have a chance to win the Arimaa challenge? Given that humans are probably thinking for longer than 2 minutes about their moves. |
|
A chance, yes, but I think not quite. I suspect the Challenge is still safe from this level of play. 7 CPU doublings should give the bots in the neighborhood of 500 rating points, plus or minus a lot. Chessandgo is still more than 500 rating points ahead, although for defenders below him it gets cloudier.
My point is that when I look back I remember thinking, "It will be difficult for a bot to do X," and now a bot can do X. X is the discovery of the past, and our current strategic understanding has moved well beyond that. Today I feel quite comfortable saying, "It is still difficult for a bot to do Y." What I am nervous about is only the future, given the past hurdles bots have overcome.
Clueless' performance in the challenge screening was 1960. Its performance in the Postal will depend on the number of games won out of 16:
Wins Performance
---- -----------
7 1945
8 2008
9 2072
10 2139
11 2210
12 2287
From the table, JDB's doubling CPU one more time to four hours per move instead of two should only give clueless one extra win. Also, humans should gain far more per doubling than clueless does, so even if humans think only, say, 16 minutes per move they are getting four doublings, which should wipe out all the gains of clueless' long thinking time. However, my intuitions and calculations about what should happen appear to be off in this instance, because they don't match the excellent tournament clueless is having. Clueless could easily end 12-4, with a score of +8 that could win the tournament!
For me thinking 16 minutes on a move is about average, but I have heard some humans (Eltripas? Tuks?) saying they think far less than that. There may be other explanations of clueless's excellent performance as well, including a statistical fluke, and including that my formulas about the value of doubling thinking time for humans and for computers are completely bogus. 
|
| « Last Edit: Jun 24th, 2010, 1:21am by Fritzlein » |
IP Logged |
|
|
|
jdb
Forum Guru
Arimaa player #214
Gender:
Posts: 682
|
|
Re: 2010 Postal Mixer
« Reply #62 on: Jun 24th, 2010, 10:54am » |
Quote Modify
|
Having clueless play in the postal mixer has been very informative for me. I have a much better idea what needs to be done to improve things. There are things that show up at postal speeds that do not show up with shorter thinking times.
About 95% of the time it is sufficient to be on the lookout for enemy goals in two. However there are times when it is required to look for enemy goals in 3. In order to have a chance at the challenge, a bot will need to be able to detect enemy goals in 3.
Once a bot has that level of goal awareness, the rest of the game boils down to trap control. Clueless got out manouvered in trap fights in several games in the postal mixer. This will be harder to fix. Adanac is very good at this aspect of the game. I think I need to improve my own understand of trap fights before clueless will make progress in this area.
|
|
IP Logged |
|
|
|
RonWeasley
Forum Guru
Harry's friend (Arimaa player #441)
Gender:
Posts: 882
|
|
Re: 2010 Postal Mixer
« Reply #63 on: Jun 25th, 2010, 5:03am » |
Quote Modify
|
on Jun 24th, 2010, 12:49am, Fritzlein wrote:
For me thinking 16 minutes on a move is about average, but I have heard some humans (Eltripas? Tuks?) saying they think far less than that. |
|
I think for far less than that, even after pondering a position for days. There must be a think/ponder ratio that is higher for some My goodness Hermione does wonders for the school uniform.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #64 on: Jun 26th, 2010, 5:38pm » |
Quote Modify
|
on Jun 25th, 2010, 5:03am, RonWeasley wrote:| I think for far less than that, even after pondering a position for days. |
|
I guess that I don't have such a high average thinking time if you don't count thinking away from the board. I am often considering what move to make at odd times, such as in the shower or walking to the bank. I would say that about half the time I don't make a move at one sitting, but let my thoughts percolate and come back to the position later. It is surprising how often I have a new perspective given a second look.
For example, I have just seen Hippo's latest move, and have considered it for about five minutes. I have a strong move candidate, and could just make the move now if I were pressed for time. I am not pressed, though, so I will go shopping instead, and think about the position as I drive. When I come back, I may see new candidate moves that I didn't before or see crushing responses for him that I didn't see before. If nothing jumps up to change my mind, I'll make the move; otherwise I'll think longer. So how you add up my total time per move definitely depends on whether you count the thinking while shopping.
I tend to think longer on games that I am losing (such as my game with Hippo), and less on games where I have the position under control. I notice from woh's rating lists that my postal ranking is 132 points ahead of chessandgo's whereas my overall ranking is 169 points behind chessandgo. I attribute this entirely to the fact that I think longer than average on my postal games.
|
| « Last Edit: Jun 26th, 2010, 5:47pm by Fritzlein » |
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #65 on: Jun 27th, 2010, 12:38am » |
Quote Modify
|
As a footnote to my opening analysis, I went back and found duplicate positions after both Gold and Silver had set up. In 144 positions, there were 138 unique positions and 6 duplicates. Half of the duplicates were variants of the 99of9 setup against each other. After 2g only three duplicates were left, and after 2s, no duplicates were left. The duplicates were
bot_clueless 722caasi
woh 722caasi
diverge 2s
starjots 722caasi
starjots willwould
diverge 2g
722caasi Weirdo87
azgreg 722caasi
diverge 2s
Heyckie ocmiente
RonWeasley Eltripas
diverge 2g
Nombril Tuks
722caasi Tuks
diverge 2g
Heyckie amalgam
Korhil amalgam
diverge 2s
This total scattering of opening setups should be enough in itself to discourage anyone from trying to codify openings, but another fact makes the situation worse. I looked up my previous research from past postal tournaments and found that opening are less standardized now than then:
In 2006, there were 3 games out of 95 that had not yet diverged after 2s. After 3g there were no duplicates.
In 2007, there were 4 games out of 85 that had not yet diverged after 2s. After 3g there were no duplicates.
So in 2010 we diverged sooner than ever.
Quote:| For comparison, last year I looked at a grandmaster chess tournament, in which there were 91 games. After two moves on each side (i.e. not counting the setup, since there is no choice) there were 20 unique positions. Many of the 71 repeats did not diverge for several more moves. |
|
Isn't that amazing? Two moves on each side (or rather the setups plus one move each) were enough to distinguish every game in this tournament, i.e. 144 out of 144. But in a chess tournament two moves on each side only led to 20 unique positions out of 91 games, i.e. 78% duplicates.
I think opening theory is one aspect in which we can just stop expecting Arimaa to be like chess.
|
|
IP Logged |
|
|
|
Hippo
Forum Guru
Arimaa player #4450
Gender:
Posts: 883
|
|
Re: 2010 Postal Mixer
« Reply #66 on: Jun 27th, 2010, 6:51am » |
Quote Modify
|
on Jun 26th, 2010, 5:38pm, Fritzlein wrote:
I tend to think longer on games that I am losing (such as my game with Hippo), and less on games where I have the position under control. I notice from woh's rating lists that my postal ranking is 132 points ahead of chessandgo's whereas my overall ranking is 169 points behind chessandgo. I attribute this entirely to the fact that I think longer than average on my postal games. |
|
OK, let it be noted you said that after 40g  I hope you just are not paying mind games ... .
I am happy to hear this after your 40 turns, but my potential lead is such small that I hardly notice it. And each turn I am losing 5 days of reserve and probably position as well .
And congrats Nombril for returning to the lead.
I suppose today (1.7.2010) clueless would lose one of the games ...
|
| « Last Edit: Jul 1st, 2010, 7:21am by Hippo » |
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #67 on: Jul 1st, 2010, 12:54pm » |
Quote Modify
|
on Jun 24th, 2010, 10:54am, jdb wrote:| About 95% of the time it is sufficient to be on the lookout for enemy goals in two. However there are times when it is required to look for enemy goals in 3. In order to have a chance at the challenge, a bot will need to be able to detect enemy goals in 3. |
|
Congrats on beating clueless, Tuks. This is a game that suggests seeing goal in three isn't enough, at least not for postal games.
http://arimaa.com/arimaa/gameroom/comments.cgi?gid=147834
I suppose, though, that I would never have the guts to sacrifice a rabbit for that game-ending goal attack, even if I had two minutes per move to think about it.
|
|
IP Logged |
|
|
|
Hippo
Forum Guru
Arimaa player #4450
Gender:
Posts: 883
|
|
Re: 2010 Postal Mixer
« Reply #68 on: Jul 1st, 2010, 1:20pm » |
Quote Modify
|
Yes congrats Tuks ... this is the game I was watching hoping for win in 3 and I was put 2 turns back in history I was scared if clueless have seen some defense I didn't ....
And grrr ... I have already made 2nd blunder against clueless (36g) ... with 73 days in reserve. Just when I start feeling I am getting out of troubles (what is the most dangerous feeling in arimaa).
|
| « Last Edit: Jul 1st, 2010, 2:33pm by Hippo » |
IP Logged |
|
|
|
jdb
Forum Guru
Arimaa player #214
Gender:
Posts: 682
|
|
Re: 2010 Postal Mixer
« Reply #69 on: Jul 3rd, 2010, 5:33am » |
Quote Modify
|
on Jul 1st, 2010, 12:54pm, Fritzlein wrote:
Congrats on beating clueless, Tuks. This is a game that suggests seeing goal in three isn't enough, at least not for postal games.
http://arimaa.com/arimaa/gameroom/comments.cgi?gid=147834
I suppose, though, that I would never have the guts to sacrifice a rabbit for that game-ending goal attack, even if I had two minutes per move to think about it.
|
|
Well done Tuks. Nicely played.
There was a goal in 4 earlier in the game. I'll leave it as an exercise for the reader to figure out where it was.
Today is the first time clueless caught up in all its postal games.
|
|
IP Logged |
|
|
|
Tuks
Forum Guru
Arimaa player #2626
Gender:
Posts: 203
|
|
Re: 2010 Postal Mixer
« Reply #70 on: Jul 3rd, 2010, 10:00am » |
Quote Modify
|
thanks guys, one of my few postal games where i had a strong position throughout.
Yes Fritz, i play postals at normal game times, the first 10 moves of each game is usually 30 seconds per move and its never much more than 2-3 minutes throughout a game unless its either really complicated or i'm losing badly
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #71 on: Jul 4th, 2010, 6:30am » |
Quote Modify
|
on Jul 3rd, 2010, 5:33am, jdb wrote:| Today is the first time clueless caught up in all its postal games. |
|
Let me ask you what I asked Janzert about OpFor: catching up gives you the option of increasing clueless' think time. Will you try to challenge humanity even more by keeping the server maxed out with longer thinks? Or will you let the server cool down and/or use it for other purposes?
I personally am glad you set it for four hours in the first place, because clueless came up with some excellent tactics that put quite a scare into me. Even now that I think I have regained control, I am treading very cautiously because clueless has proven it deserves more respect than I was giving it at first.
|
|
IP Logged |
|
|
|
jdb
Forum Guru
Arimaa player #214
Gender:
Posts: 682
|
|
Re: 2010 Postal Mixer
« Reply #72 on: Jul 5th, 2010, 5:38am » |
Quote Modify
|
on Jul 4th, 2010, 6:30am, Fritzlein wrote:
Let me ask you what I asked Janzert about OpFor: catching up gives you the option of increasing clueless' think time. Will you try to challenge humanity even more by keeping the server maxed out with longer thinks? Or will you let the server cool down and/or use it for other purposes?
I personally am glad you set it for four hours in the first place, because clueless came up with some excellent tactics that put quite a scare into me. Even now that I think I have regained control, I am treading very cautiously because clueless has proven it deserves more respect than I was giving it at first. |
|
I need my computer for other things, so it will stay at 4 hours per move.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #73 on: Jul 5th, 2010, 12:01pm » |
Quote Modify
|
on Jul 5th, 2010, 5:38am, jdb wrote:| I need my computer for other things, so it will stay at 4 hours per move. |
|
Thanks for having devoted that server for so long already so that clueless could play in the Postal Mixer. It has been a treat for me, and I think for others as well.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: 2010 Postal Mixer
« Reply #74 on: Jul 9th, 2010, 1:40am » |
Quote Modify
|
on Jun 23rd, 2010, 8:22am, jdb wrote:| I have the log file for clueless for most of the postal tournament. The first couple weeks or so are missing. If someone reminds me, I'll post it after the games are all finished. |
|
I'd be interested to see clueless' evaluations throughout my game, in particular around move 19 when clueless threw away its considerable advantage. How long was it before the bot knew it had messed up?
I congratulate clueless for tactically outplaying me on a couple of crucial moves. After that, however, my opinion of clueless' ability was inflated so much that I was shocked at its helplessness the rest of the game. Clueless was completely incapable of cutting its losses or stopping the bleeding in any way. In contrast, I realized after the fact that my 15g was a blunder, and on 16g I gave up the camel hostage I had been holding to try to limit the damage.
Clueless saw its forced loss on move 32, I believe, but between move 19 and move 31 its evaluation must have deteriorated by about a rabbit per move. That's what I call going down in flames. I'd like to see that plot of evaluation versus time to see whether it was a steady slide to oblivion, or case of clueless remaining in denial for many moves before falling off a cliff. Either way I'm not sure how it would be fixed.
Funny side note: Clueless pulled my rabbit into hostage position at a6 on move 17s and didn't have time to take it until move 35s as a "spite capture" just before I goaled.
http://arimaa.com/arimaa/gameroom/comments.cgi?gid=148577
|
|
IP Logged |
|
|
|
|
Home
Help
Search
Members
Login
Register
Reply
Notify of replies
Send Topic
Print