--> Turn 8 it's strange that the bot accepts getting pushed out of the Scandinavian Peninsula.

Perhaps it prioritized eliminating the enemy from East US (worth 5) higher than preventing elimination from Scand (worth 3)? That's exactly what happened in the game.

https://www.warlight.net/MultiPlayer?GameID=10627518
--> Very strange expansion turn 1, this doesn't seem right.

Does it have some sort of priority to have a stack built up? The way it expanded early was kinda defensive, by completing the bonus with a stack rather than minimal troops.

Perhaps it prioritized eliminating the enemy from East US (worth 5) higher than preventing elimination from Scand (worth 3)? That's exactly what happened in the game.

Well, I know how the bot is supposed to behave. Scandinavia immediately gives the opponent a bonus when the bot gets eliminated there so the Java bot would have focused there.


Does it have some sort of priority to have a stack built up? The way it expanded early was kinda defensive, by completing the bonus with a stack rather than minimal troops.
It should have taken the center territory to be capable to complete the bonus next turn. This never caused any problems with the Java bot.

From this, am I getting that the plan is to work on Wunderwaffle rather than start from scratch?

I don't think so, Fleece. I think that Wunderwaffe is just the first example of an AI other than Prod that got donated to Warlight.AI under open source.

The only thing I worry about is if we all work on different AIs, we'll end up doing more work for the same result. Unless you're just intending to make an AI for just yourself, of course...

Hmm, I was thinking that things would start with a bunch of different AIs, then pick the most promising looking one from the bunch and work on that. You're right there will probably be duplicated effort that way, but I'm not sure that's avoidable for something like this AI project. Who knows what the best kind of AI will be?

I suppose that's true...

i thinkai should have set levels so you could have it as good as a level 60 player and stuff like that

If the AI gets too good, getting booted (and having the AI take over in a team game) might be advantageous. Currently, there is a large gap between the strength of the AI and how strong the AI would have to be before booting your teammates became a viable strategy. If you reduced this gap by about half, it would be good. If you reduce the gap too much, it would be a problem... you want getting booted to be a bad thing.

This argues for having different levels of AI available, especially if you guys come up with a very strong AI.

Any AI would never be able to do as well as a good human in a team game, simply because the AI can't communicate with their teammates.

i still think ais should pick terrtorys after the team mates

How does this thing have a bootrate?

How does this thing have a bootrate?
- Code crashes
- The orders get send from the computer of the starting player, he can abort the AI calculations

I am currently investigating the cowzow AI, written by Dan Zou, so that we can have a third bot to play with. Looking at the leaderboards:



We're fortunate enough all of the top 3 bots on this leaderboard have donated their source code to WarLight: GreenTea, Wunderwaffe, and Cowzow. Wunderwaffe is the one already available in the framework.

I originally started working on the GreenTea bot, however I ran into some technical problems with it which led to me switching to Wunderwaffe.

So far, the results of Cowzow look promising. It placed higher than Wunderwaffe on the leaderboard, runs much more quickly, but best of all it has a much smaller code footprint. Wunderwaffe is over 10k lines of code, where Cowzow is under 3k. This makes it easier to work with. (GreenTea weighs in at a hefty 15k. The current production AI is simple, at under 1k).

Once I get it adapted to WarLight.AI I'll add it to the github repo.

^ Great!

PS: I did knew you needed a lot of code lines to do an AI, but 15k is really insane lol.

@TeamGuns: For me under 3000 seems very little. There is also some standard stuff like the data structures, the game engine communication,.... Then also all AIs need some heuristics about what is going on behind the fog, some basic calculations like for shortest paths,... I guess in Wunderwaffe some minor helper stuff that I more or less wrote multiple times could get compressed however 3000 lines of code is really little.

The reason Wunderwaffe is pretty fast is that by design it doesn't look into future configurations and as a current flaw the "tasks" get executed sequential instead of weighing them. I'm wondering how cowzow can beat this speed, perhaps it's because Wunderwaffe calculates stuff quite often without reusing old results.

I believe that the assignment of territory values is still broken what was actually the "Wunderwaffe" of this bot, knowing exactly where it hurts his opponent most (the ported bot attacks the opponent in pretty much random places). However some other stuff is starting to work again and the play is taking form:

https://www.warlight.net/MultiPlayer?GameID=10662034

I would be very curious to see how any of these AIs do on my negative bonus map. The current warlight AI chokes on it.

https://www.warlight.net/Map/14287-Obama-Economy

That negative bonus map would be difficult to program around (That is assuming that the AI recognises the bonuses, and either does not expand because on average all territories have a negative bonus value, or expands, and runs into the problem of having a negative income).
A02 fell for the same sort of thing in this game: https://www.warlight.net/MultiPlayer?GameID=9923408

It should be pretty easy to get the AI to interact with negatives correctly, just need to get the formula for bonus calculation correct. The real trick will be getting the AI to recognise when a negative bonus is worth taking for positional advantage...

It would also be worth figuring out how to make it so that the AI takes all but one from the negative, while still having the gain from the positive.