This blog is intended to be a location where feedback can be given from the community regarding my Diplomacy AIs....
Cheers,
Jason van Hal
Albert Specific Comments
I encourage specific criticisms. Keep it civil. Albert is no longer a work in progress. Maybe it is a hobby I will pick up again in the future. I hope others get enjoyment out of it for personal play, and that eventually it will help to train a better AI.
Version 5.4t, which I just played against, had some trouble at the end...
France (bot) had a random fleet floating around in the Barents and couldn't kill of a three-dot Germany with English help in time to keep Turkey from soloing.
Yes it definitely is tough deciding when it is safe to take a small power out that close to endgame. Albert is a little cautious, figuring it is better to not always try and whittle down the draw at all costs.
I have had the opportunity to test 1 human vs 6 Albert v 5.8 in like 15 level 20 press-games. I am impressed with the challenge they give me. In the end game I have found it hard to form a stalemate line with Albert(s) against a soloing power, though. And they are annoyingly honest with replies to peace/alliance propositions. But I have been stabbed more than once by false responses to my propositions, so I assume there is some intelligence for the replies. Keep the good work up!
Thanks. I have already reworked the deceit levels for responses. I had a bug that went unnoticed. I am sure it will be a while before I get the right balance of honesty. In the next version, each Albert will start with a predisposition level of deceit. Eventually I plan that this level will get adjusted depending on a powers situation.
And even harder to beat... Sweet. But I can't get Albert to understand any move/support proposals. What am I doing wrong? Is there any documentation for the communication Albert understands at level 20/30 press?
this is simply terrible. the game is a good trainer, however, as your opponents make better moves that you didnt think possible. This project is simply flawed, however, in how the alberts act and getting them to mimic players. Their opening moves are unrealistic whatsoever and often i see alberts at each other's throats, and then best buddies for no reason at all
your efforts could be put use on something else, on diplomacy is a waste of time
With such a keen eye, I hope you will try and write an AI that can compete. And when you do let me know the name. The DAIDE project allows everyone to try and build a better mouse trap. Good Luck.
Any chance a future update can allow the # of SCs for solo victory be set at launch? I keep finding myself one to three points away from victory, while Albert is one and I'm turning the tide. I end up losing by a hair simply because he can exploit other powers' inability to defend themselves. :(
I actually haven't tried the AI diplomacy piece. I've just been reading the moves the AI has been telegraphing and it works fairly well if you're defensive. Another thing I noticed ... it seems whether I set the difficulty to 0 or 100, the AI is still difficult. Kudos for such a magnificent feat! Any chance this will be ported to Android? ;)
Thanks for the comments Andrew. Sorry the number of SCs for victory is part of the DAIDE protocol that Albert is designed to use.. and it can't be changed.
As for the difficulty... it is a bit of a misnomer. The setting should really be called 'depth of thought' The higher the number the more iterations Albert will perform. As it turns out, Alberts running at higher values will on average beat those at lower values. Whether this is noticeable when playing a human is not sure. You should notice a bit of a difference, especially if the setting is less than 10.
So in short I never really spent a lot of time trying to make in less smart, just gave it a throttle on how much time to spend thinking.
Sorry no plans for porting it. I hope to release a version that has decent press in it.
Just got into diplomacy and discovered this. You spoke of a potential update in Sept 2011. Is it likely be to released soon or is it still quite a ways off?
I am thinking in the next month. It is basically done, I just haven't found time and hasn't been on the top of my list. But if I get a few requests, it would encourage me to do so.
Hmm. I have installed it on Windows 7 machines before. I can post it as a zip. But I'm wondering if you have a virus scanner that is giving problems. Can you disable that while you install.
Very disappointed that linux doesn't seem to be available. The disappointment would be less if your front page was accurate. Please change it to say "Windows only", as linux and other operating systems run on PCs, so the game is far from "PC only".
At least that way I wouldn't have wasted as much time on it.
Just getting back into Diplomacy after a long time of not playing. This has let me play so many games. Playing Albert 6 Gunboat I note that: I seem to win every time as Turkey Albert's end game is not definitive enough. One game lasted with three powers (two Alberts and me) until 1933. It appeared that unless I was the largest the two Alberts would always attack the largest player and not necessarily the one from who they could win the most SCs and possibly take victory. Would love to be able to specify the starting year and season and unit positions and then release the game for 7 Alberts to play 1000 games for the next few moves...
You can if you get into editing the Daide files. Check out the DipAI forum. The first part is easy enough. Setting up the map in a specific year as the starting point. I had an observbot that records the outcome of the games to a CSV file.
But I agree. Albert is programmed to attempt to stop the power it thinks is closest to winning. It was a safety first programming. Prior to this Albert would too often stab for a few centers, not make the victory, and piss of his alliance so they threw the game. It is difficult to determine these situations, and Albert doesn't have those smarts.
The way I left Albert was that if he played as a single player with 6 humans, he would not do something stupid to mess the end game when there was a clear objective of stopping the leader. Whether I succeeded or not, I don't know.
I just wanted to say thanks for making Albert. I used to play Diplomacy many years ago in high school and have enjoyed playing it recently. It plays a good game.
I just played a few oppening games (not finished) and I'm already impressed. I'll be glad to share more impressions when I have time to properly test your AI
Anyway, I have some considerations to make, but im not sure if its the proper forum.
as for specific comments, I have two questions to make:
1. I read your comment in 2012 saying "The bots can't communicate with each other". Is it still true? If it is, why not? And how does it make plans and use eachothers support if they dont communicate?
2. does it have layers of deceifullness (great concept, by the way!)? I mean, when I play, even when I choose to be deceitfull and go for solo from the beggining, I try to convince my opponents that I'm trustfull. There are many ways to determine when to stab or when to lie, based on mutual perception of deceitfullness. I havent played enought to say its a flaw or something, just wondering
1) The bots can communicate in a game with press, just not in a no-press or gunboat game. The comment was made to assure the player that there was no illegal communication happening. No collusion in a gunboat game.
2) I wouldn't say it is fantastic at deceitfulness. It basically makes decisions through the course of the game when might be an opportune moment to stab. It does not pre-plan it from the beginning. It may decide a few turns before it actually stabs that it was going to do so, but other priorities restricted it from doing so.
Albert is good at tactics, but it is no strategic genius. I tried to give it many of the elements needed, but there were limitations within the framework. My real intention for Albert was to make a good benchmark AI that more serious AIs could train against, which hasn't really happened yet. It is good for humans to play against to dust the cobwebs off.
are you still interested in improving it? Im not complaning, it already feels like a great achievment. If I were paradox, I'd definetely want to hire you
I would like to play with 5 Albert's difficulty 100 with 2 human players via DAIDE Mapper, I can adjust difficulty with QuickStart Diplomancy but this is only single player. How can I play 2 humans and adjust difficulty?
Hi Jason, I'm an artificial intelligence researcher building a diplomacy bot and looking to benchmark it against Albert. I'm having difficulty running Albert, as it seems to only be distributed as Windows executable, and I don't have access to a Windows machine. I wonder if you'd be willing to share the source code so that I could compile it for Linux? Would love to hear from you at jonathan.sam.gray@gmail.com
Is it possible to enter an existing spring or fall board into Albert and have it propose "best" moves? Or is history to significant a part of the AI engine?
Version 5.4t, which I just played against, had some trouble at the end...
ReplyDeleteFrance (bot) had a random fleet floating around in the Barents and couldn't kill of a three-dot Germany with English help in time to keep Turkey from soloing.
Yes it definitely is tough deciding when it is safe to take a small power out that close to endgame. Albert is a little cautious, figuring it is better to not always try and whittle down the draw at all costs.
ReplyDeleteThis comment has been removed by the author.
ReplyDeleteVersion 5.8 is now released
ReplyDeleteI have had the opportunity to test 1 human vs 6 Albert v 5.8 in like 15 level 20 press-games. I am impressed with the challenge they give me. In the end game I have found it hard to form a stalemate line with Albert(s) against a soloing power, though. And they are annoyingly honest with replies to peace/alliance propositions. But I have been stabbed more than once by false responses to my propositions, so I assume there is some intelligence for the replies.
ReplyDeleteKeep the good work up!
Thanks. I have already reworked the deceit levels for responses. I had a bug that went unnoticed. I am sure it will be a while before I get the right balance of honesty. In the next version, each Albert will start with a predisposition level of deceit. Eventually I plan that this level will get adjusted depending on a powers situation.
ReplyDeleteVersion 5.9 now released
ReplyDeleteAnd even harder to beat... Sweet.
ReplyDeleteBut I can't get Albert to understand any move/support proposals. What am I doing wrong?
Is there any documentation for the communication Albert understands at level 20/30 press?
No, Albert is not capable of move orders yet.
ReplyDeleteSee the webpage for some of its capabilities:
http://sites.google.com/site/diplomacyai/albert/coding-press
this is simply terrible. the game is a good trainer, however, as your opponents make better moves that you didnt think possible. This project is simply flawed, however, in how the alberts act and getting them to mimic players. Their opening moves are unrealistic whatsoever and often i see alberts at each other's throats, and then best buddies for no reason at all
ReplyDeleteyour efforts could be put use on something else, on diplomacy is a waste of time
With such a keen eye, I hope you will try and write an AI that can compete. And when you do let me know the name. The DAIDE project allows everyone to try and build a better mouse trap. Good Luck.
ReplyDeleteAny chance a future update can allow the # of SCs for solo victory be set at launch? I keep finding myself one to three points away from victory, while Albert is one and I'm turning the tide. I end up losing by a hair simply because he can exploit other powers' inability to defend themselves. :(
ReplyDeleteI actually haven't tried the AI diplomacy piece. I've just been reading the moves the AI has been telegraphing and it works fairly well if you're defensive. Another thing I noticed ... it seems whether I set the difficulty to 0 or 100, the AI is still difficult. Kudos for such a magnificent feat! Any chance this will be ported to Android? ;)
Thanks for the comments Andrew. Sorry the number of SCs for victory is part of the DAIDE protocol that Albert is designed to use.. and it can't be changed.
ReplyDeleteAs for the difficulty... it is a bit of a misnomer. The setting should really be called 'depth of thought' The higher the number the more iterations Albert will perform. As it turns out, Alberts running at higher values will on average beat those at lower values. Whether this is noticeable when playing a human is not sure. You should notice a bit of a difference, especially if the setting is less than 10.
So in short I never really spent a lot of time trying to make in less smart, just gave it a throttle on how much time to spend thinking.
Sorry no plans for porting it. I hope to release a version that has decent press in it.
Hey,
ReplyDeleteJust got into diplomacy and discovered this. You spoke of a potential update in Sept 2011. Is it likely be to released soon or is it still quite a ways off?
Thanks,
I am thinking in the next month. It is basically done, I just haven't found time and hasn't been on the top of my list. But if I get a few requests, it would encourage me to do so.
ReplyDeleteI get an error when trying to get the Albert 6.0 installer to work. I downloaded the zip files for 5.5 and they work fine.
ReplyDeleteWhat operating system are you using?
ReplyDeleteWindows 7 Home Premium.
ReplyDeleteHmm. I have installed it on Windows 7 machines before. I can post it as a zip. But I'm wondering if you have a virus scanner that is giving problems. Can you disable that while you install.
ReplyDeleteYup... that was it. I turned off Norton and it installed. Thanks for your help! Keep up the good work on Albert!
DeleteThis AI taught me how to play diplomacy. I love it.
ReplyDeleteVery disappointed that linux doesn't seem to be available. The disappointment would be less if your front page was accurate. Please change it to say "Windows only", as linux and other operating systems run on PCs, so the game is far from "PC only".
ReplyDeleteAt least that way I wouldn't have wasted as much time on it.
Just getting back into Diplomacy after a long time of not playing. This has let me play so many games. Playing Albert 6 Gunboat I note that:
ReplyDeleteI seem to win every time as Turkey
Albert's end game is not definitive enough. One game lasted with three powers (two Alberts and me) until 1933. It appeared that unless I was the largest the two Alberts would always attack the largest player and not necessarily the one from who they could win the most SCs and possibly take victory.
Would love to be able to specify the starting year and season and unit positions and then release the game for 7 Alberts to play 1000 games for the next few moves...
Thanks for the fun.
You can if you get into editing the Daide files. Check out the DipAI forum.
DeleteThe first part is easy enough. Setting up the map in a specific year as the starting point.
I had an observbot that records the outcome of the games to a CSV file.
But I agree. Albert is programmed to attempt to stop the power it thinks is closest to winning. It was a safety first programming. Prior to this Albert would too often stab for a few centers, not make the victory, and piss of his alliance so they threw the game. It is difficult to determine these situations, and Albert doesn't have those smarts.
The way I left Albert was that if he played as a single player with 6 humans, he would not do something stupid to mess the end game when there was a clear objective of stopping the leader. Whether I succeeded or not, I don't know.
I just wanted to say thanks for making Albert. I used to play Diplomacy many years ago in high school and have enjoyed playing it recently. It plays a good game.
ReplyDeleteThanks for that!
ReplyDeleteI just played a few oppening games (not finished) and I'm already impressed.
I'll be glad to share more impressions when I have time to properly test your AI
Anyway, I have some considerations to make, but im not sure if its the proper forum.
thanks!
as for specific comments, I have two questions to make:
ReplyDelete1. I read your comment in 2012 saying "The bots can't communicate with each other". Is it still true? If it is, why not? And how does it make plans and use eachothers support if they dont communicate?
2. does it have layers of deceifullness (great concept, by the way!)? I mean, when I play, even when I choose to be deceitfull and go for solo from the beggining, I try to convince my opponents that I'm trustfull. There are many ways to determine when to stab or when to lie, based on mutual perception of deceitfullness. I havent played enought to say its a flaw or something, just wondering
congrats!
1) The bots can communicate in a game with press, just not in a no-press or gunboat game. The comment was made to assure the player that there was no illegal communication happening. No collusion in a gunboat game.
ReplyDelete2) I wouldn't say it is fantastic at deceitfulness. It basically makes decisions through the course of the game when might be an opportune moment to stab. It does not pre-plan it from the beginning. It may decide a few turns before it actually stabs that it was going to do so, but other priorities restricted it from doing so.
Albert is good at tactics, but it is no strategic genius. I tried to give it many of the elements needed, but there were limitations within the framework.
My real intention for Albert was to make a good benchmark AI that more serious AIs could train against, which hasn't really happened yet. It is good for humans to play against to dust the cobwebs off.
thanks for your quick reply
Deleteare you still interested in improving it?
Im not complaning, it already feels like a great achievment. If I were paradox, I'd definetely want to hire you
I would like to play with 5 Albert's difficulty 100 with 2 human players via DAIDE Mapper, I can adjust difficulty with QuickStart Diplomancy but this is only single player. How can I play 2 humans and adjust difficulty?
ReplyDeleteHi Jason, I'm an artificial intelligence researcher building a diplomacy bot and looking to benchmark it against Albert. I'm having difficulty running Albert, as it seems to only be distributed as Windows executable, and I don't have access to a Windows machine. I wonder if you'd be willing to share the source code so that I could compile it for Linux? Would love to hear from you at jonathan.sam.gray@gmail.com
ReplyDeleteIs it possible to enter an existing spring or fall board into Albert and have it propose "best" moves? Or is history to significant a part of the AI engine?
ReplyDelete