Α.Ι.

A forum to seperate the more serious discussions from the lighter topics in Off-topic.

Re: Α.Ι.

Postby WHSeward » 27 Sep 2016, 01:05

Noefochesgiving wrote:Is there an AI that plays a decent Gunboat game?

No.

Diplomacy is a much harder game to solve than Checkers (solved - optimal play is always a draw) or Chess (an excellent game can be played, but it is not solved.) There are two big reasons for this.

1) The traditional approach for solving games is brute force; do a comprehensive search of board positions and choose the best one given your current position. That is how Checkers was solved. Chess has too many positions to search them all, so they have to be clever and do a deep search down a limited number of options, until the end-game when there are fewer pieces on the board and a comprehensive search is possible.

But in Diplomacy the search space is far, far too large to handle this way. While the absolute number of positions on the board is smaller than Chess, since all the pieces move at once, the number of available positions in just a few plies is enormously larger. It just isn't practical to do a brute force search of positions.

2) Then there the problem of evaluating board positions. While not trivial, it is reasonably straight forward to say whether one position is better than another in Chess. Not so in Dip. Because Dip is a game of cooperation, the best tactical position in Dip is not necessarily the best position to help you win. The fact that Dip strategy cannot be reduced to Dip tactics, again violates how we go about solving games today.

Bottom line: sure Dip is analyzable but it will take totally different programming techniques than what we have used for other games.

Edit: none of this by the way is "artificial intelligence." Brute force techniques are sort of the opposite of A.I. and that is state of the art for games today.
"As a general truth, communities prosper and flourish, or droop and decline, in just the degree that they practice or neglect to practice the primary duties of justice and humanity." WHS

A member of the Classicists.

Ask me about mentor games. Send me a PM or post in the Mentoring forum.
User avatar
WHSeward
Premium Member
 
Posts: 2912
Joined: 29 Dec 2012, 22:16
Location: San Francisco, California, USA
Class: Star Ambassador
Standard rating: (1633)
All-game rating: (1647)
Timezone: GMT-8

Re: Α.Ι.

Postby super_dipsy » 27 Sep 2016, 07:09

Actually I am fairly convinced it is possible for a gunboat Diplomacy AI today. It is tempting to try to build one ;) . Even full Diplomacy should be possible given some time, at least to build a pretty strong player (although not a guaranteed winner, there is not enough source data available).

While WHS is right that there are many combinations, given all pieces move, modern AIs generally don't work like this. Yes, Big Blue cracked Kasparov at Chess by doing the analysis of billions of possible outcomes, but if you look at the way Watson won Jeopardy that was a totally different problem. That was to do with language parsing and a horrendously wide general knowledge. It is all about probabilities and weightings and likelihoods.

Watson components are now available in the Cloud, and can be used to assemble applications as required. The point about Watson (only picking on this specific because it is probably the best developed) is that it has a much wider way of looking at things, and more particularly it LEARNS. This is how it won Jeopardy. Exactly the same technique could enable it to be all-powerful at gunboat Diplomacy at least. I would hazard a guess that it could aspire to winning normal Diplomacy too; all that is required is the programs to be put in place and then give it time to learn.

Perhaps a real life example of the Watson-like approach to AI might be helpful, because it also shows one of the major limitations of the technology today. One of the first applications of Watson was in Oncology. Some healthcare providers use it as an assistant to the Oncology expert. It takes as input the patient notes, results of tests carried out and the worldwide sources on cancer to a) recommend additional tests and more importantly b) recommend the possible actions options the specialist can take with in a weighted list, with associated likelihoods of success. The limitation of course is that at this stage we humans are not ready to trust this sort of decision totally to an AI engine, and perhaps more importantly the specialist can take into account things that Watson is unaware of such as the mood of the patient today. But in fact the specialists who use Watson assign great weight to its recommendations, combining them with their own skills to choose the preferred action. Actually, the more fascinating (although less critical) example I like is the use of the technology to predict flu outbreaks; in this application, it analyses Twitter and Facebook and other social media to check for keywords and sentiment to map potential flu outbreak paths!

So given input of Diplomacy games, all the communications, results etc and then using its tools such as sentiment analysis, a Watson-based Diplomacy engine would get better and better to the point that it could reliably win. Yes of course it would have some level of 'possible combination' analysis, but the point is it does not rely on this type of brute force attack of working out every possibility but instead on qualitative information, for example to be found in language used in messages and outcome likelihoods.

For anyone interested, take a look at IBM's Bluemix at http://www.bluemix.com and in particular the Watson section of the catalog. These are all cloud-based components you can snap into your applications to build your own AI!
User avatar
super_dipsy
Site Admin
Site Admin
 
Posts: 11517
Joined: 04 Nov 2009, 17:43
Class: Ambassador
Standard rating: (1000)
All-game rating: (956)
Timezone: GMT

Re: Α.Ι.

Postby WHSeward » 27 Sep 2016, 07:35

Yeah that is right, Dipsy, it will take that kind of approach to try to crack Dip, a real A.I. approach. I am just not people trying to do that with Dip or in the game space generally. (I do know there has been some great progress on Poker lately and I don't know the approach that has been used, so I could be wrong. But then, Poker is much simpler analytically.)

Setting Poker aside, it seems to me that even though every now and then we read about a new game that has been solved, it is just sort this esoteric exercise because it is always another massive search approach. It seems a pretty limited application technique. A self-learning approach is going to be what it takes for Dip.
"As a general truth, communities prosper and flourish, or droop and decline, in just the degree that they practice or neglect to practice the primary duties of justice and humanity." WHS

A member of the Classicists.

Ask me about mentor games. Send me a PM or post in the Mentoring forum.
User avatar
WHSeward
Premium Member
 
Posts: 2912
Joined: 29 Dec 2012, 22:16
Location: San Francisco, California, USA
Class: Star Ambassador
Standard rating: (1633)
All-game rating: (1647)
Timezone: GMT-8

Re: Α.Ι.

Postby Carebear » 27 Sep 2016, 15:40

I have not played Gunboat Diplomacy on this site, so I am not familiar with the various options for public communication. However, I have played it in person and on the old eMail server Judge back in the late 80s and early 90s. IIRC, the Press options were No Press, White Press, Grey Press, and Black Press, with various combinations of the coloured.

I agree that a Gunboat Diplomacy AI is doable and should be able to be pretty strong, for No Press. I think it could be equally strong for White Press, but I would have some reservations about implied knowledge even in White Press.

I think that Black Press and Grey Press add a dimension that would require another level of capability that does not currently exist. As I under stand it, there were concerns about the framing of questions in Jeopardy and other AI limitations in the Watson challenge. Watson had the buzzer advantage. Any question it didn't answer was likely because the phrasings approached these limitations. Any communication that involves embedded/implied additional information I believe will require capabilities yet to be developed.

A Darwinian development about lying can be found here. However, this is simplistic and non-AI oriented.

More recent developments about lying in communication can be found here and here.

Dipsy,

Many online Go sites publish service information to allow robots to participate in games. Though an interface could be written to spoof a browser, perhaps a feature could be developed for the new system where a listener is available with a public service description for robots to connect to play. Access through this port could have limitations to gunboat only games. Of course, then we would probably want Gunboat games to have an option during creation as to whether Robots are allowed to join.

A successful Gunboat Diplomacy AI would be a great first milestone to developing a full Diplomacy AI (and then the end of human existence).
You can have my last supply center, when you pry it from my cold dead hands.

Spam Ambassador Wannabe

Officially Sanctioned Site Gadfly (meaning the negative kind of sanction)
User avatar
Carebear
 
Posts: 2389
Joined: 12 Nov 2013, 04:26
Location: In the fingerhold
Class: Star Ambassador
Standard rating: (1573)
All-game rating: (1589)
Timezone: GMT+8

Re: Α.Ι.

Postby WHSeward » 28 Sep 2016, 00:30

There was an effort to run a tournament for DIPbots last year, the Computer Diplomacy Challenge. The winning entry came from a post-doc in Australia (a Netherlands native.) He got a paper out of it, but my sense is that this was not an earth-shaking bot.

I don't think they got enough entrants either. They didn't repeat the event this year as far as I am aware.
"As a general truth, communities prosper and flourish, or droop and decline, in just the degree that they practice or neglect to practice the primary duties of justice and humanity." WHS

A member of the Classicists.

Ask me about mentor games. Send me a PM or post in the Mentoring forum.
User avatar
WHSeward
Premium Member
 
Posts: 2912
Joined: 29 Dec 2012, 22:16
Location: San Francisco, California, USA
Class: Star Ambassador
Standard rating: (1633)
All-game rating: (1647)
Timezone: GMT-8

Re: Α.Ι.

Postby Strategus » 29 Sep 2016, 21:00

Carebear wrote:More recent developments about lying in communication can be found here and here.

I love the idea of a "Bolloxometer". Can we prioritise the development of that first?
The Devil makes work for idle forces

Better to have fought and lost, than never to have fought at all
Actual Platinum Classicist
I did WDC 2017

Just say "NO!" To carebears and kittens
User avatar
Strategus
 
Posts: 1570
Joined: 30 May 2015, 14:30
Location: England
Class: Star Ambassador
Standard rating: 1659
All-game rating: 1721
Timezone: GMT

Re: Α.Ι.

Postby Carebear » 29 Sep 2016, 23:21

GPD wrote:
Carebear wrote:More recent developments about lying in communication can be found here and here.

I love the idea of a "Bolloxometer". Can we prioritise the development of that first?

Really don't want to see that available to my opponents.
You can have my last supply center, when you pry it from my cold dead hands.

Spam Ambassador Wannabe

Officially Sanctioned Site Gadfly (meaning the negative kind of sanction)
User avatar
Carebear
 
Posts: 2389
Joined: 12 Nov 2013, 04:26
Location: In the fingerhold
Class: Star Ambassador
Standard rating: (1573)
All-game rating: (1589)
Timezone: GMT+8

Re: Α.Ι.

Postby Bromley86 » 11 Oct 2016, 05:29

I haven't tried it yet, but I saw this:

https://sites.google.com/site/diplomacyai/home
A member of the Classicists, a group that aims to reduce NMRs/surrenders.
Bromley86
 
Posts: 503
Joined: 02 May 2012, 00:16
Class: Star Ambassador
Standard rating: (1880)
All-game rating: (2233)
Timezone: GMT+12

Re: Α.Ι.

Postby DoranMartell » 11 Oct 2016, 14:53

Bromley86 wrote:I haven't tried it yet, but I saw this:

https://sites.google.com/site/diplomacyai/home


I just tried it a few times. It works pretty good. In any case, it's way better than the Avalon Hill game (but that should not come as a surprise). So I guess it's at least good enough to practise some tactics as a starter in Diplomacy.
DoranMartell
 
Posts: 16
Joined: 13 Nov 2015, 23:40
Class: Star Ambassador
Standard rating: (1179)
All-game rating: (1180)
Timezone: GMT

Re: Α.Ι.

Postby Noefochesgiving » 03 Nov 2016, 18:32

Bromley86 wrote:I haven't tried it yet, but I saw this:

https://sites.google.com/site/diplomacyai/home



Well I tried playing it no press and I have to say I was challenged. The AI is better than the average Gunboat player I have played against here. Would be a great addition to Gunboat PD when you can't get a full roster or there is a surrender.
User avatar
Noefochesgiving
 
Posts: 58
Joined: 19 Jun 2016, 17:36
Location: The Border region between Upper and Lower Canada
Class: Diplomat
Standard rating: (992)
All-game rating: (1171)
Timezone: GMT-5

PreviousNext

Return to Debates

Who is online

Users browsing this forum: No registered users and 1 guest