There has been some talk of WarLight being a candidate for the next Google AI Challenge. For those not familiar with the Google AI Challenge, the GAIC is a competition where software engineers write a bot to play a game for them. They then upload their bot into the GAIC website which matches their AI against everyone else’s. It’s a lot of fun – they’ve featured other popular games such as Galcon and the Tron game, which is a bit like the old snake games.
To prepare for WarLight as a candidate, I threw together a prototype of a framework that can run a WarLight AI Challenge. This contains everything you need to write and run a WarLight AI on your own computer, completely separate from the WarLight servers.
I appreciate any feedback on this prototype – shoot me an email at email@example.com!
WarLight is a very flexible engine that supports many different variations and settings. For this prototype, I selected settings that I think would make for a good AI challenge. I chose these settings both to make the game as fair as possible, as well as making it as easy as possible to write an AI.
I chose the Double Earth map for this prototype, since it’s symmetrical and therefore can be completely fair to both parties. For the initial territory layout, I pick three random territories from the top earth and give these to one AI, and give the same three from the lower world to the other AI. The rest of the territories start out as neutral 2s.
Most strategic WarLight games let players choose where they start. This is entirely plausible for the AI Challenge as well, however I decided that a random (but mirroed) approach was better simply because it makes it easier to write an AI. Having to write code to determine where to start is a fun AI challenge itself, but doing this adds more syntax to the AI language and is just one more thing you need to learn and write before you can have a working AI. If the GAIC sponsors want this turned back on it can be.
For those unfamiliar, WarLight contains a dozen different types of cards that have a variety of effects on the game. I decided for the AI Challenge that cards should be left out, mostly because they add more syntax to the AI language and, like above, increase the base knowledge you need to learn to write an AI.
Fog is the WarLight feature that controls whether or not you can see the entire map or just what’s next to you. Whether or not to include fog in an AI challenge is an interesting decision that I thought a lot about. Ultimately I decided to turn it on, since it’s a feature unlike things that the GAIC has done before. If you’d like to try it without fog, you can turn it off in the downloaded source on line 34 that says “useFog: true”
WarLight’s rules do contain some randomness, which enhances the game experience for human players. However, for an AI challenge, it may be beneficial to remove the use of the random number generator. There are four places that Warlight uses the RNG: distribution, combat, rounding, and move order. Two of these have already been removed from the prototype, and the other two could also be eliminated if we wanted to get rid of the RNG completely.
– Distribution: As mentioned above, bots don’t get to choose where they start so the normal randomness that affects strategic 1 v 1 games during distribution has already been removed for this AI challenge.
– Combat: WarLight’s main combat rules contain some use of the RNG. WarLight already contains a slider that allows you to control how much you want the RNG to affect your game. For the AI challenge, I have set this slider to 0, which eliminates this use of the RNG. If you want to turn it back on, you can do so in the downloaded source on line 33 that says “luckModifier: 0”
– Rounding: WarLight uses random rounding, where 0.4 has a 40% chance of being rounded up and a 60% chance of being rounded down. This is left turned on in the prototype, however if the GAIC sponsors wish, we could change this to a straight round to eliminate this use of the RNG.
– Move order: WarLight randomizes the order that players move within a turn. Although this does have an effect on the game, it’s typically not the deciding factor. Nevertheless, if the GAIC sponsors wish, we can change this to order based on a predictable heuristic, such as alternating turns or giving first move to the player with fewer armies or something like this.
How to Write an AI
For those unfamiliar with GAIC, this prototype uses standard in/out to communicate with bots. This means you can write your bot in any language that can print to and read from the console. The fastest way to get started is to take a look at the C# starter pack included in the downloadable below as it contains a fully functional (although not incredibly smart) AI.
At Startup: The Map
WarLight maps are obviously a graph. When your AI is started, the first thing it will do is read the map from standard-in. The first line is simply the word “Map” informing you that a map is coming up.
After that, your AI will read a series of lines starting with TerritoryDefinition. For example:
TerritoryDefinition 1 SolomonIslandsN 48 8 TerritoryDefinition 2 St.PetersburgN 87 4 11 110 111 104 TerritoryDefinition 3 PeruN 125 96 147 64 TerritoryDefinition 4 PermN 110 92 2 87 70 TerritoryDefinition 5 MyanmarN 26 29 49 119 TerritoryDefinition 6 PanamaN 147 25 TerritoryDefinition 7 ShanghaiN 50 26 151 114
Territory definition arguments:
– Argument 0: The word TerritoryDefinition
– Argument 1: The ID of this territory
– Argument 2: The name of the territory with spaces stripped out. This is not used by your AI, but it assists in debugging as knowing the name of the territory makes it easier to locate on a map.
– Argument 3+: The IDs of the territories that this territory connects to.
After reading territories, your AI will be fed a series of lines starting with BonusDefinition. For example:
BonusDefinition 5 CentralRussiaN 83 45 120 105 75 137 150 BonusDefinition 4 AustraliaN 27 122 14 135 134 12 91 15 BonusDefinition 6 EastRussiaN 136 115 66 18 46 103 65 129 72 149 BonusDefinition 6 EuropeN 54 51 109 34 10 80 11 101 95 BonusDefinition 4 MiddleEastN 138 52 21 53 43 19
Bonus definition arguments:
– Argument 0: The word BonusDefinition
– Argument 1: The number of armies this bonus is worth if you control all of its territories
– Argument 2: The name of the bonus. This is not used by your AI, but it assists in debugging as knowing the name of the bonus makes it easier to locate on a map.
– Argument 3+: The IDs of the territories that make up this bonus.
After the bonuses, your AI will be fed a line that says “EndMap” and nothing else. This notifies you to stop reading the map and to get ready to start the first turn of the game.
Each Turn: The Standing
In your main game loop, first read a line that will say either “GameOver” or “Turn #” where # is the turn number. Upon reading GameOver, you should obviously exit your process.
If you don’t see GameOver, you should start receiving a game “standing.” A standing is the state of the board at one point in time – it tells you who controls each territory and how many armies on are on it.
Territory 198 ownedBy Neutral withArmies 1 Territory 199 ownedBy Neutral withArmies 2 Territory 200 ownedBy You withArmies 5 Territory 209 ownedBy Neutral withArmies 2 Territory 214 ownedBy You withArmies 5
– Argument 0: The word “Territory”
– Argument 1: The territory ID
– Argument 2: The word “ownedBy”
– Argument 3: Either “You”, “Neutral” or “Opponent”
– Argument 4: The word “withArmies”
– Argument 5: The number of armies on this territory
Keep reading territories until you read the word “EndStanding” alone on a line.
Each Turn: Send Your Orders
After reading the standing, it’s time to execute your custom AI logic! As we’ve removed cards, the only thing your AI needs to issue are Deploy orders and AttackTransfer orders, which is what can order armies to attack or transfer from one territory to an adjacent one.
You should print out orders to standard out like this:
Deploy 4 on 5 Deploy 1 on 112 AttackTransfer 4 from 5 to 29 AttackTransfer 3 from 5 to 119 AttackTransfer 8 from 26 to 119 end
– Argument 0: The word “Deploy”
– Argument 1: The number of armies you’re deploying
– Argument 2: The word “on”
– Argument 3: The territory ID you’re deploying on
– Argument 0: The word “AttackTransfer”
– Argument 1: The number of armies to attack or transfer
– Argument 2: The word “from”
– Argument 3: The territory ID to attack or transfer from
– Argument 4: The word “to”
– Argument 5: The territory ID to attack or transfer to
Finally, after issuing all of your orders, simply print out “end” to signal that you’re done. That’s all there is to it!
How to run your AI
The downloadable below contains an executable named WarLightAIChallenge.exe. Simply run this with two arguments that point at the bots you wish to execute. For example, to have the included starter pack fight itself on Windows, run: (note: Windows users must have .net 4.0 installed)
WarLightAIChallenge.exe StarterPack.exe StarerPack.exe
On OSX or Linux, run it in Mono 2.8 or higher:
mono WarLightAIChallenge.exe StarterPack.exe StarerPack.exe
For this prototype, I have not yet written a visualizer. This means that there’s no provided app that lets you view your game data in a graphical format to easily see what’s going on. If anyone writes one and is willing to share, send me an e-mail and I’ll post it here.
This package contains:
– StarterPack source code
– StarterPack pre-compiled binary
– AI runner source code (named WarLightAIChallenge)
– AI runner pre-compiled binary (named WarLightAIChallenge.exe)
If you run into any problems or have any comments, please send me an email at firstname.lastname@example.org.
6 thoughts on “WarLight as a candidate to the Google AI Challenge”
I think WarLight is even better than ants, but there is no turning back for closest contest. Thanks for adapting engine and starter package. Waiting for visualizer and for next AI contest.
Yeah, this is aimed at the next contest after Ants.
I liked the AI challenge last year, and I like Warlight.
The combination of these would be awesome!
It definietly worth participating.
Unfortunately, I do not have time for testing now.
Whoa. No, really, that’s surprising. I’m definitely taking part in this (even though I was too lazy to participate in previous Challenges).
Won’t it be possible to use attack/transfer-only orders, though?
Attack-only and transfer-only aren’t included in this prototype just to make it easier to write a bot.
They could easily be added though. Just look at line 264 of the engine, there’s a parameter that says AttackTransferEnum.AttackTransfer. This can specify .AttackOnly or .TransferOnly – you can parse it out of the line your bot sends right here.
Report from trying this on Ubuntu 11.10 with Mono 2.10.5: The programs did not work out of the zip file as is. I had to change the executable permission for the sample bot so that the WarLightAIChallenge.exe was able to run a test game. Before this change it only reported a fatal exception about failing to start a process.
So, in my case this was necessary: ‘ chmod u+x StarterPack.exe ‘ to make the bot executable.