In the article Roles of Artificial Intelligence in the S2006M issue of the DP Zine, Edi Birsan set out his view of the way that the development of Diplomacy AIs should progress in order to best benefit the hobby. This article compares his ideas with the current state of Diplomacy AI development.

Edi introduces the article with a quote on the current state of AI development, primarily the DAIDE project. Then, after describing the problems with the e-mail hobby, he says, "What is needed is for us to have a modified AI that fits in with the e-mail play of the game, not one that is designed to play against a single player in the classic computer game".

Contrary to what this sentence implies, the DAIDE project is not designed for single player play. Each power is played by a separate player; that player can be a human or one of several different AIs that have been developed. This allows games to be played in any configuration. The "classic computer game format" of one human and six AI players is played regularly by various people, especially those involved in the DAIDE project who want to see how their AI is playing. We've also played games across the Internet, and across local networks, with two humans and five AIs, right up to six humans and one AI, and also games with 7 humans and no AIs at all, or 7 AIs and no humans at all. The system handles all of these configurations equally well. Where there are multiple AIs (or multiple humans) in the same game, powers have no information as to which other powers are humans and which are AIs.

The bulk of the text in Edi's article is based around the statement that, "We need the AI to act as a backup for the player to keep his position viable and to have the possibility to act to take over abandoned positions to preserve the game flow. Here is an outline of what might be a step-by-step for the backup AI." Edi then continues by proposing a step-by-step method which could be used to generate a set of orders for a player who has been NMRed. This step-by-step approach has a number of serious flaws in it.

Step one is: "Player creates a profile for his position: allies, enemies, neutrals, priority of builds and removals." This is meant to be an AI which is looking after a position for a player who has NMRed. If the player has time to create a profile for his position, then surely he has time to perform the much less time-consuming task of entering some orders. If he doesn't, then it's going to have to be done by the GM, who should not be making judgement calls as to what the alliance structure on the board is, etc.

The step-by-step guide then goes on to detail how to generate a set of default moves, builds, and disbands. The obvious question here is, if such a simple algorithm could really provide a much more suitable set of orders than an NMR "all hold", then why has a GM never run a game with such an algorithm used to generate moves for powers which NMR? After all, the algorithm he suggests is so simple that a GM could work out the orders it generates in a minute or two, using just pencil and paper. The only reasons I can think of are either that it doesn't work (that the moves are sometimes much better, but sometimes much worse, than all units holding), or that players want games where NMRs are punished. Either way, the evidence suggests that the method Edi suggests isn't what people actually want.

Furthermore, if it were possible to write an effective AI with such a simple algorithm, then we would have had competitive Diplomacy AIs years ago. Unfortunately, it's not that easy.

Edi then goes on to say, "Once we have the AI able to step in for missed adjustments, and retreats, and then one-off situations, then we can move to it being able to stand in for taking over abandoned positions for a few turns at a clip". The fact is we already have such an AI. Any of the current AIs can do both these things. It just takes a little lateral thinking.

The AIs are designed to play any variant that has been coded up. Therefore, an easy way to generate a set of orders for a single turn is to create a variant where the starting position is the current position in your game, and see what opening moves the AI comes up with. To do this, you have to create a variant within the DAIDE system.

A DAIDE variant is coded in four files. The first two are the most technical and complicated. The symbol file defines which symbols represent provinces in this variant, which represent powers, etc, and how they are encoded into the DAIDE system. The map file defines the map; which provinces are adjacent to which others, which provinces are land, sea, or coast, which are supply centres, etc. But the good news is, for a variant that is just a different starting position, these are unchanged — so you can just copy the files for the Standard variant.

Thirdly, there is the current centre ownership file. It looks like this:

SCO (AUS BUD TRI VIE) (ENG LVP EDI LON) (FRA BRE MAR PAR) (GER KIE BER MUN)
(ITA ROM NAP VEN) (RUS STP MOS WAR SEV) (TUR ANK CON SMY)
(UNO NWY SWE DEN HOL BEL SPA POR TUN GRE SER RUM BUL)

And fourthly, there is the unit position file. It looks like this:

NOW (SPR 1901) (AUS FLT TRI) (AUS AMY VIE) (AUS AMY BUD)
(ENG FLT LON) (ENG FLT EDI) (ENG AMY LVP)
(FRA FLT BRE) (FRA AMY PAR) (FRA AMY MAR)
(GER FLT KIE) (GER AMY BER) (GER AMY MUN)
(ITA FLT NAP) (ITA AMY ROM) (ITA AMY VEN)
(RUS FLT (STP SCS)) (RUS FLT SEV) (RUS AMY MOS) (RUS AMY WAR)
(TUR FLT ANK) (TUR AMY CON) (TUR AMY SMY)

Neither of these should be too hard to modify to enter the current position of a game. The only catch is each province has one abbreviation, and only one. For instance, you must use ECH for English Channel, and MAO for the Mid Atlantic.

Once that is done, run up the server, select your new variant, run up seven AIs, let them play, and then have a look in the log and see what they ordered for the first turn. And there you have your default orders.

Edi then goes on to say, "Finally, once the above is done, then and only then do we move toward the creation of the AI as a player making proposals and what not, starting with a basic format of who is an ally, enemy and neutral to working on specific combinations with other powers to play."

The problem with this approach, and one of the fundamental policies that the DAIDE project has had from the start, is that nobody knows how to write a good Diplomacy AI. If you don't allow people to write negotiating AIs until we have good no-press AIs, then you are making the unfounded assumption that the best AIs will use their tactical and strategic ability to work out what to say, rather than using their negotiating abilities to work out what to order. If that turns out to be wrong, then we would have wasted years of development time on strategy-first AIs, before starting to consider negotiation. That is why the DAIDE project has, from the start, had a design that allows press and no-press play between AIs.

Finally, Edi says, "This is at least a stepladder approach that will allow the hobby to have some benefit to each of the simple small steps in the progression of an AI than waiting for someone to come up and create a total AI package from the start."

"The total AI package" has been around for over three years now. DumbBot was released in April 2003, and was the first AI to win a game with all the other 3 powers being played by humans (it was only a 4-player variant). The more recent AIs are much better than DumbBot, although, as far as we are aware, no AI has yet won a game of Standard Diplomacy against 6 human players.

If you want to try playing against the AIs, then go to www.ellought.demon.co.uk/dipai . Download the DAIDE Server, DAIDE Mapper and DumbBot. And then click the link at the top of the page to see the instructions on how to use them. Once you're beating DumbBot on a regular basis, go to http://www.daide.org.uk/clients.xml and download the latest AIs — especially KissMyBot and R.E.D., and see how you fare against those.

If you'd rather play against a mixture of humans and AIs, then the best place to find a game is on the Yahoo Groups mailing lists RTDip and RTNow. Games are regularly arranged on there, with anything between zero and five AIs in the game, depending on how many people are around. The environment is so quick to use that we've been known to have a no-press game play to 1913 in about 45 minutes!

All in all, the AIs are already around, to take part in games in whatever way you want. And if you think you can do better, then join us on the DipAI mailing list on Yahoo Groups and have a go at developing your own.

So, to summarise, if you read Edi's article and thought that it would be really good if AIs could ever get that far — well, they're already doing everything that Edi asked of them, and a whole lot more.

David Norman (david@ellought.demon.co.uk)

If you wish to e-mail feedback on this article to the author, and clicking on the envelope above does not work for you, feel free to use the "Dear DP..." mail interface.