Author |
Topic: Postal Tournament (Read 6764 times) |
|
Fritzlein
Forum Guru
Arimaa player #706
Gender: 
Posts: 5928
|
 |
Re: Postal Tournament
« Reply #60 on: Jan 30th, 2005, 10:48am » |
 Quote  Modify
|
on Jan 29th, 2005, 10:18am, fotland wrote:| How much time are the the strong players using per move in this contest? I bet it's under 10 minutes per move. |
|
I find I'm averaging around 10 minutes per move. If school gets busier, that may become less. However, I notice that most of my opponents slowed down from the initial blistering pace, which takes some of the pressure off, and allows me to think 10 minutes per game without feeling rushed.
Simply waiting for faster computers won't be enough, since humans are still rapidly getting better. You would have to improve Bomb as fast as humans are improving AND have computers get faster if you are going to gain ground.
It's no surprise that Bomb thinks it is ahead in most games. In my game I too think the position is objectively much in favor of Bomb. I agree with Bomb's evaluation. But the reason I played four rabbits forward and the elephant/horse attack is that I don't think Bomb will play well in such a position. Even if Bomb wins, it leaves open the question of what would happen if I tried to play the objectively strongest move rather than anti-Bomb moves. But if Bomb loses, that will show a huge remaining gap, i.e. it will show that extra search depth doesn't compensate for Bomb not understanding strategically how to beat off the elephant/horse attack.
For example, on Bomb's last move it could have made a threat to frame my horse that would have been very tricky to counter. Instead it moved its second horse *away* from the contested trap, which meant I didn't have to counter the frame threat and could clear my exposed rabbit from the center. Bomb will face another strategy test on this move: it can again set up a frame threat which is tricky for me to answer. However, I fully expect Bomb won't do this, as part of a pattern of not making the strongest move in this type of position. (After Bomb moves, I'll edit this post to either gloat or eat crow.  )
[Edit] Looks like I get to gloat, at least over move 7b. Moving the cats away too makes a frame very remote. One would have to conclude that if Bomb can't see a way to immediately force a frame *tactically*, then Bomb has no *strategic* interest in threatening one.
|
| « Last Edit: Jan 31st, 2005, 7:22am by Fritzlein » |
 IP Logged |
|
|
|
fotland
Forum Guru
Arimaa player #211
Gender:
Posts: 216
|
|
Re: Postal Tournament
« Reply #61 on: Jan 30th, 2005, 12:41pm » |
Quote Modify
|
I haven't given Bomb an evaluation that favors threats that can be countered, even if the coutner is tricky. Against people it is good to set up tricky moves, but the search algorithm assumes the opponent will see everything bomb can see.
I don't see how you can avoid the horse frame, but bomb didn't find a forced horse frame last move. However, it did think you had to defend rather than move your rabbit, so we'll see
I thought bomb was very clever the way it set up the horse frame in Naveed's game. I want to see if it is smart enough to rotate out its elephant.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Postal Tournament
« Reply #62 on: Jan 31st, 2005, 7:39am » |
Quote Modify
|
on Jan 30th, 2005, 12:41pm, fotland wrote:I haven't given Bomb an evaluation that favors threats that can be countered, even if the coutner is tricky. Against people it is good to set up tricky moves, but the search algorithm assumes the opponent will see everything bomb can see.
|
|
Ah, this isn't what I had in mind, although it would be interesting to attempt. I guess that since Omar provides the opponent name as part of the game state, you could set up Bomb to make the "objectively best" move against other computers, and to make "anti-human" moves against humans at fast time controls in hopes of a blunder.
But anyway what I was referring to was keeping up the pressure. The way I like to attack (as a fallible human) is usually just to put pressure on in a way that limits the defensive choices and forces my opponent to make certain contortions in defending. Often I can't see any concrete gain at the end of my lines, but I still think "This is a position where something has got to give eventually", or "There must be a way for me to get something out of this line, even though I don't see it now."
When Bomb fails to pressure me in this way, I can make my own plans at leisure. At the moment (move 8w) I have only to worry vaguely about Bomb's potential camel attack on the west wing, and am otherwise at leisure to proceed with my strategic aims of the opening: either getting Bomb's small pieces offsides where I threaten them, or making a general rabbit advance on the east wing to threaten remote goal, constrain Bomb's elephant, threaten to liberate my elephant, etc. We'll see whether Bomb at this time control creates enough counterplay to stop me; speedy usually doesn't.
|
|
IP Logged |
|
|
|
99of9
Forum Guru
Gnobby's creator (player #314)
 
Gender:
Posts: 1413
|
|
Re: Postal Tournament
« Reply #63 on: Jan 31st, 2005, 4:09pm » |
Quote Modify
|
on Jan 26th, 2005, 1:45pm, RonWeasley wrote:OK, 8 then. And that's my final offer.  |
|
Well as far as I can tell you just executed a 36 ply forced goal against me!
I'm back in Australia now. Omar could you change my nationality in the player database? Thanks.
|
|
IP Logged |
|
|
|
fotland
Forum Guru
Arimaa player #211
Gender:
Posts: 216
|
|
Re: Postal Tournament
« Reply #64 on: Jan 31st, 2005, 5:30pm » |
Quote Modify
|
I'm not convinced that an AI breakthrough is needed to build a strong Arimaa program. Bomb uses pretty standard computer chess techniques (PVS, iterative deepening, transposition table, iterative deepening, null move, singular extension, killer and history heuristics, etc.), and it's pretty strong already.
More importantly, the evaluation function and search extensions are far from optimal, and I'm still getting big strength improvements for not much work.
The ultimate bottleneck is more likely to be access to advice from a strong player, since I'm not a very strong Arimaa player myself.
|
|
IP Logged |
|
|
|
Keith
Forum Junior Member
Arimaa player #4
Gender:
Posts: 9
|
|
Re: Postal Tournament
« Reply #65 on: Feb 1st, 2005, 4:57pm » |
Quote Modify
|
[quote author=fotland link=board=talk;num=1104134643;start=45#58 date=01/29/05 at 10:18:37]How much time are the the strong players using per move in this contest? I bet it's under 10 minutes per move. If so, Bomb has a 20 to 1 time advantage compared to the world championship.
I can't speak for the strong players. In order to juggle ten games in the time I have available I am taking 2 to 5 minutes per turn averaging about 3 minutes.
|
|
IP Logged |
|
|
|
99of9
Forum Guru
Gnobby's creator (player #314)
Gender:
Posts: 1413
|
|
Re: Postal Tournament
« Reply #66 on: Feb 2nd, 2005, 3:19pm » |
Quote Modify
|
My game against Mr Brain has been reset (and it seems all his other postal tourney ones have disappeared). He was down to about 5 days of thinking time on his second move. Should I start again??
|
|
IP Logged |
|
|
|
MrBrain
Forum Guru
Arimaa player #344
Gender:
Posts: 148
|
|
Re: Postal Tournament
« Reply #67 on: Feb 4th, 2005, 3:27pm » |
Quote Modify
|
Wow, what happened to all of my games? I was very busy and unavailable until now. I came in to make moves in all of my games, but they have disappeared. Please advise!
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Postal Tournament
« Reply #68 on: Feb 4th, 2005, 6:32pm » |
Quote Modify
|
on Feb 4th, 2005, 3:27pm, MrBrain wrote:| Wow, what happened to all of my games? I was very busy and unavailable until now. I came in to make moves in all of my games, but they have disappeared. Please advise! |
|
I don't know whether the moves that already happened can be recovered, but even if they can't at the very least the games should be restarted with three or four days in reserve for you. There must be something special in the eyes of the server about games where no moves have been made except for the initial setup of pieces.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Postal Tournament
« Reply #69 on: Feb 5th, 2005, 8:52pm » |
Quote Modify
|
I ran a spreadsheet on the predicted outcome of the postal tournament, assuming that the ratings at the start of the tournament were accurate. Of course the ratings were not accurate for any number of reasons, but if they were, here is the expected number of points each player should score out of their ten games.
99of9 2171 8.47
Fritzlein 2164 8.42
Belbo 1933 6.07
Omar 1918 5.98
Bomb 1880 5.69
RonWeasley 1851 5.53
Naveed 1810 5.34
Adanac 1780 5.23
Paul 1687 4.56
Robinson 1658 4.42
MrBrain 1605 4.19
Junaid 1570 3.87
Jdb 1556 3.93
Keith 1481 3.04
Kamikazeking 1471 2.96
Rabbitball 1412 2.3
The numbers I'm most curious to compare with the actual results are MrBrain's and Bomb's. MrBrain has long agitated for slower time controls in part because he expects to do better at them, so if he's right he should score vastly better than 4.19, and in fact should have a good shot at winning the whole tournament.
Bomb's score will be interesting because Bomb is actually thinking for three hours per move compared to a few minutes per move for the rest of us. This would suggest a score much higher than 5.69, and indeed also contend for the tournament victory. On the other hand, it is my intuition that computers just aren't very good at long time controls, so I wouldn't be surprised if the prediction of the ratings is accurate for Bomb despite the huge thinking time differential.
|
| « Last Edit: Feb 5th, 2005, 8:56pm by Fritzlein » |
IP Logged |
|
|
|
99of9
Forum Guru
Gnobby's creator (player #314)
Gender:
Posts: 1413
|
|
Re: Postal Tournament
« Reply #70 on: Feb 5th, 2005, 9:14pm » |
Quote Modify
|
Interesting numbers Fritz.
I agree that I'll be looking closely at bot_Bomb and MrBrain, but also at RonWeasley, who hadn't previously played humans. So far he's made a great start of it - beating me, and in a close battle with Belbo. For that matter, Belbo's results will be interesting too, given that he's the human representative this year. Although I guess that's a bit unfair because he'll be putting a lot of effort into defending the title at the same time.
For my part, I don't think I have a hope of getting 8.47. I'm definitely better at short time controls where I can rely on my opponents to overlook my plans . And anyway, I've already lost 1... so can only afford a draw from here on in. A lot of my games look like they're going to be very close battles. One very interesting one is my game with Paul, which until a move or two ago had settled into a quasi-equilibrium where neither of us could go for much advantage because both had a horse on either wing.
In fact I think everyone will be nearer to 5.0 than predicted ... my feeling is that without blunders, everyone becomes much more equal in arimaa. But we shall see - perhaps i'm just being optimistic.
|
|
IP Logged |
|
|
|
MrBrain
Forum Guru
Arimaa player #344
Gender:
Posts: 148
|
|
Re: Postal Tournament
« Reply #71 on: Feb 6th, 2005, 11:55am » |
Quote Modify
|
It would be nice to have the current games and results in a cross table in the Postal Tournament page. Looking at a list of 71 current games doesn't really give anyone a sense of how the games and tournament are progressing.
|
|
IP Logged |
|
|
|
Fritzlein
Forum Guru
Arimaa player #706
Gender:
Posts: 5928
|
|
Re: Postal Tournament
« Reply #72 on: Feb 6th, 2005, 2:01pm » |
Quote Modify
|
on Jan 31st, 2005, 5:30pm, fotland wrote:I'm not convinced that an AI breakthrough is needed to build a strong Arimaa program. Bomb uses pretty standard computer chess techniques (PVS, iterative deepening, transposition table, iterative deepening, null move, singular extension, killer and history heuristics, etc.), and it's pretty strong already.
|
|
In the early days of chess programming, folks had sharply different intuitions as to what it would take to get a comptuer to play chess well. Arguing about it didn't seem to change anyone's mind; what settled it was the brute force searching camp winning lots of games. Similarly Omar's dreams of AI breakthroughs could be crushed by good engineering working well in pratice for Arimaa. I really don't know what to expect: If jdb posted tomorrow and said he got clueless quite strong primarily by AI techniques that wouldn't surprise me either.
I agree that Bomb is quite strong already compared to the level of human play right now. Yet I don't think you have gained any ground from a year ago, since humans are still learning so much about the game so quickly. The ratings gap between Bomb and the top-rated human is about 300 points, just like last year. I expect that next year you will again have to sweat just to keep Bomb within 300 points of the top-rated human.
on Jan 31st, 2005, 5:30pm, fotland wrote:More importantly, the evaluation function and search extensions are far from optimal, and I'm still getting big strength improvements for not much work.
|
|
It was quite palpable how much Bomb improved due to your work this winter. It made it much more fun and interesting to play speedy when some old tricks stopped working. But the areas you list for further improvement (the evaluation function and selecting lines for extended search) are just the sort of things that AI people would love to claim that their techniques are particularly good at.
I wonder if you will soon feel that you hit a wall in being able to improve the hard-coded evaluation.
on Jan 31st, 2005, 5:30pm, fotland wrote:The ultimate bottleneck is more likely to be access to advice from a strong player, since I'm not a very strong Arimaa player myself.
|
|
Ah, but how useful is the advice that I or other players give? I can say that Bomb's last move in our postal game was useless -- just shifting rabbits along the back rank, if anything away from where they needed to be. How would you stop Bomb from making that move next time? Or when I say Bomb should have threatened to frame my horse to keep up the pressure, the idea of "pressure" isn't useful to you. Yet I know for sure that if Bomb just sits there holding my horse hostage with his elephant and threatening nothing else, I will squeeze him to death eventually. It's waiting for me to blunder, and if I don't, I will win.
The most concrete advice I can give to improve Bomb's strategy at the moment is that, when faced with an elephant/horse attack, it MUST frame the attacking horse, or be doomed to long-term disadvantage. It's great if one can bring about the frame without using both one's elephant and camel, but usually they are both necessary. The camel has to come help out if the elephant doesn't have enough friends nearby to do the job.
I am coming to believe that Bomb's current idea of attacking with its camel when it holds a horse hostage is simply inadequate. It doesn't create enough counterplay, particularly when there is no time pressure to make humans blunder tactically. Also, the camel can't operate in the opposing trap near the opposing elephant, and since there is one opposing trap it can't contest, it loses a trickle of cats and rabbits and dogs into that trap. Plugging that leak would help, but I don't think it can substitute for the necessity of framing the hostage horse.
I hope that helps you improve the evaluation function, because I really would like to see Bomb improve and stop me from winning in the same old way. My anti-speedy strategy seems to be working just fine against the postal Bomb that can think for three hours, so a faster chip is probably not the answer.
|
|
IP Logged |
|
|
|
fotland
Forum Guru
Arimaa player #211
Gender:
Posts: 216
|
|
Re: Postal Tournament
« Reply #73 on: Feb 6th, 2005, 10:19pm » |
Quote Modify
|
The games are far enough along now that it is becoming clear that more thinking time can't overcome Bomb's evaluation deficiencies. Even Bomb thinks that Fritz, Omar, and Paul are up about a rabbit each, and Belbo is up 0.5 rabbit. It only looks to me to have winning positions in two games.
It's passed up several chances to frame pieces or rabbits, so I need a lot more work there. It doesn't understand that you need more than one piece to attack a trap, so that is another obvious improvement. It doesn't undertand a near-hostage, where an advanced weak piece can't retreat, but isn't actually a hostage yet. It doesn't understand that framed doesn't stop a second piece from being hostage at the same time. These are the next evaluation enhancements I'll do.
I'm sure there are also many bugs in the evaluaiton function to be fixed. One problem with deeper search is that it is more likely to find a way to get to position that an infrequent evaluation bug thinks is good.
One way you could help would be to help me identify missed opportunities. For example in the bomb-belbo postal game, at 23b, it looks to me that Bomb could immobilize Belbo's Elephant by moving its Elephant to b5 and rabbit to c7. Later it could pull one of Belbo's rabbits forward to release its own Elephant, and have a winning position. Is there something I'm missing that makes this a bad plan? If not, it shows a big problem with the Elephant mobility evaluation.
|
|
IP Logged |
|
|
|
Paul
Forum Junior Member
Arimaa player #888
Gender:
Posts: 8
|
|
Re: Postal Tournament
« Reply #74 on: Feb 7th, 2005, 8:13am » |
Quote Modify
|
Some toughts :
Bomb evaluates his position as inferior in 4 games. In 3 of them, he plays silver.
I think I can put them in two categories : Omar and Fritz games in on side, and Belbo and mine in the other.
Both Omar and Fritz seem to have found a position where Bomb does not know what to play. It has given a rabbit without any counter-play in his game against Omar, and has the same problem in 9b against Fritz : advancing the cat from e8 to e7 is a terrible mistake, clearly made because Bomb has no ideas what to play.
On the other side, both Belbo and I have tried to use at our advantage Bomb setup (this is of course true too for the two other games, but I try to keep things simple).
Putting the camel on the far wing of the opposite elephant when playing silver has led him to a difficult position.
In my game, his phant has to defend both f3 and f6 traps.
His camel has almost no activity so far.
Belbo has choosen to attack directly the camel, so he can pull a rabbit to the c3 trap, and free the other wing for his horse.
I think it's the first reason of the relative weakness of Bomb : to avoid repetitive games, David has randomized the setup. The problem is that bomb do not know what to do whith some of them. It's a big difference between humans and bots : humans begin to play on move 1, choosing a steup where they feel comfortable, and having already some ideas about the type of game they will play. Bomb begins to think on move 2.
In fact, the same problem appears in the middle of the game : Bomb is clearly the reactive player : he does not take initiative, but punish bad moves/planifications of the opponent. When confronted to good defense or flat situation, it simply plays repetitive moves.
Now, I have to say that i'm putting a lot of thinking and time in my game against bomb. I take often 20 to 30 minutes to play my turn, and calculate some sharp variants to a depht of 20-24 plys.
It's the only game of the ten where I spend so much time, and by far.
At this subject, 10 games is way to much in my opinion to play them all at good level.
|
|
IP Logged |
|
|
|
|
Home
Help
Search
Members
Login
Register
Reply
Notify of replies
Send Topic
Print