Warlight-AI was thrown in by Fizzer into the competition

there was also some talk from jim and fizzer that the winner of the challenge being turned into a warlight single player challenge for humans to try and beat the bot on those settings.

Hey Greentea, nice job on running away with all the ELO-points :D (they will soon all be his) I'm wondering if you're still just improving your general model or if you made some specific changes for the map?

Hello

Nah, Blender is the most badass bot at present and will probably take number 1 spot.

One mistake however:
http://theaigames.com/competitions/warlight-ai-challenge/games/53a44cbc4b5ab20b42223fe0
http://theaigames.com/competitions/warlight-ai-challenge/games/53a43eb14b5ab20b42223ee6
--> Blender isn't expanding effectively in Europe. It's better to attack neutrals with one big attack than with two small attacks.

Maybe you / someone else finds a mathematical / algorithmic solution to effectively expand. This solution has to satisfy following constraints / find a good compromise:
1. The expected chance to get the bonus has to be >= the chance to get the bonus with each other expansion plan (using the same amount of deployment)
2. After taking the bonus the leftovers should not be in the middle of nowhere but where they are needed.
3. While taking the bonus also the leftovers have to be in a place where you can use them for the next expansion step (belongs to condition 1).
4. If it's unlikely that it will help you taking the bonus faster you shouldn't split your stack.
5. Sometimes it's better to hit a territory strong which borders the opponent than going for a crazy stack splitting. There are countless times where my bot won in Africa and then lost the game since he tried to take all Africa spots in one turn while the opponent sat in South America. I'd say this is always the case if the opponent bot knows you are in that bonus and also the case if your chance of taking the bonus is low anyways.

Ikalgo, just fix some bugs xD
But you (Blender, right?) will be soon #1 - imba Europe strategy works.
Such staregy should be very easily to code, because it is well splitted on states, with performing best moves in each state:

1 -> Capturing Europe - all is deployed here
Success 1:
2 -> Attack strong to North Africa
Success 2:
3.1 -> Concentrate forces in North Africa, attack if needed, and in same time go to Australia.
go to step 5.
Failure 2:
3.2 ->
If Europe is broken. go to step 4.
If Enemy in North Africa has much less armies then W.Europe + S.Europe go to step 2.
If Armies in W.Europe < Armies S.Europe - deploy all in W.Europe and move from S.Europe to W.Europe. go to step 3.2
If Armies in S.Europe < Armies W.Europe - deploy all in S.Europe and move from W.Europe to S.Europe. go to step 3.2
Failure 1:
4 -> Loss :(
5 -> Win :)

Edited 6/20/2014 18:02:21

Hey,

Yea, ironically it was just fixing one bug (twice) from the version that was competing (which I mentioned a couple of times in the forum during the finals), and removing my SA strategy alltogether.

The perfect expansion of Europe is indeed quite complex. I agree with all your points made and I have considered all of them early on as this was vital to my europe strategy. I wasn't able to solve "graph theory" for it.

I was also considering to just use a game tree which considers all possibilities but I haven't really found out how to do it. The constraint of ending up in the southern regions in the end gave me a headache more than once.

You noticed the suboptimal choices in the "last turns of capturing Europe" but I found them not that problematic most of the time (though, indeed it is suboptimal), as at that late stage of capturing europe, e.g. attacking with 7v2 is only marginally better than 6v2 (and thus the split is not that harmful). Add in the possibility that an enemy is near and it's not that hard to have a suboptimal solution occur.


My heuristics are like the following:

Placement:
Basically I try to aim for each region to have (neutral_troop_count + 2) for all the regions in Europe. This you can also still see in that last turn of capturing Europe where it would be more efficient to place them all on 1 big stack as you point out, but "finishing up" a continent is a bit different than "optimizing chance to capture a continent a.s.a.p. (i.e. expansion within continent)". Indeed, a better algorithm would combine them. I guess I could add "finishable" state or something and then run another algorithm to make a big stack.

Attacking:
- make [neutral, my_region] pairs of neutrals in europe and my regions that can attack them
- sort the pairs by how many of my regions can attack the neutral region (fewer first) and as second condition by the number of European neutral neighbours they have (fewer first)
- make an empty object that will store which of my regions have already attacked
- for each of those pairs, if my region of the pair has not attacked yet, attack the first occurring neutral with (neutral_troop_count + 2), add region to "done" object.

The result is usually a good ordering of attacks

- new loop where I loop over my regions and allow them to attack another region if they have enough troops
- new loop where I loop again over my regions and have them attack the region which has lowest probability of being captured (if they have enough troops).

But yea, I'd like to see someone solving the solution as it for sure is possible to obtain a perfect strategy. It is quite a weird tree with it's dynamic nature (e.g. there are quite a huge amount of states for a game in "round 3" if you consider all the different numbers of troop count with 100% luck divided over a variable amount of regions owned, plus the fact that Europe is asymmetrical giving many different options).

I do believe that this algorithm will be extremely important in the next competition.

Edited 6/20/2014 18:19:27

GreenTea: You basically describe what I did :-). And yea, my main issue was that I was afraid it wouldn't work and therefore I also coded SA strategy, which in my case was not as good as you guys' solution. Btw, it would be nice if your bot would lose some and let me be first for a bit so I'm first in both Poker and Warlight ;D

The thing that made Europe work in the end was choosing a region in SA. That really slows you down. I added this a day before final, but I hadn't anticipated that it would mean my troops would arrive quickly on North Africa, and then either run into your big stack that was used for capturing SA, or my troops would even run away if encountering an uncaptured brazil (none of this happened in my earlier version where I did not choose a dummy region in SA. the issue with a dummy in SA was that Dalek was coming for me from Australia at a high speed. at that time I still tried to capture Brazil rather than just settle for North Africa. The correct order is: north africa->indonesia->brazil->greenland)! I now fixed that simply and it's working great.

I chose the name Ikalgo here as it is the name of an octopus in a show that I like. I see the Europe strategy as an octopus; Europe is the body, it extends his tentacles to NorthAfrica/Brazil, Australia and to North America ;) It has to mainly protect his body, but also the endings of his tentacle haha.

Main struggles:

- Protecting Europe TwoVOne (I now implemented something that tries to counter you guys)
- Expanding in Europe

Edited 6/20/2014 20:49:34

Already looking forward to season 2... random maps should be great fun!

Now you guys can feel my pain. Will it blend, YES!




PS. someone update their bot to pick Europe spots plskthx.

Yea, it's funny that I haven't seen a Europe bot in the whole competition... Kewakareban is the closest to it; he sometimes captured Europe quite fast.

It's a great day for me -- first time in warlight on #1, and at the same time Poker #1 :)

Edited 6/22/2014 11:18:06

I seem to remember getting beat early in the qualifiers by SunTzu with a Europe strategy, he didn't do it half as well though and failed to climb in the ranks. Gave me some early loses though before I climbed out of his range. Then I got paired up with Blender twice and had no chance. And now he's refined it to dominate everyone until someone can update to counter-pick him and toss him out of his comfort zone. Not sure what he does when others take a spot in Europe. He's dominating right now though. Currently has a 15 win streak.

The point is that people can take a region in Europe, I will fallback to my old SA tactics (if I own less than 2 regions in Europe). It would certainly be an upgrade in odds, but that would also mean that those who counter-pick will have less chance against ALL other bots, since why would you want your choice to be in Europe against other bots, if not playing Europe yourself?

There are better ways to deal with Europe :-)

Edited 6/22/2014 15:22:04

Europe is a weak strategy that doesen't need to be explicitely countered since it should always lose when the opponent plays proper warlight gameplay ;)

The problem is that at present there aren't any bots with a next generation map understanding. However when the opponent bot takes Australia and he sees you not deploying in South America then he needs to find out under which circumstances he will lose. North America / Asia will give him an auto win. Also Africa is worse than his position if he takes South America. So to not lose this won position he has to go strong into North Africa before fighting in South America / taking South America. This seem to be to difficult to implement without hardcoding but I don't really think so. Even without exploiting this special map you need to realise that this is a small fighting map and have to write your code accordingly. Without stalemate situations you aren't supposed to expand at all and you need to crush the opponent instead of building your own empire.

I agree that it can be countered. That's why I chose to play 50-50 SA/EU tactics in the first place, since I wasn't trusting EU anymore (seems like that was a mistake for now). Considering that next-gen won't be around the corner (especially for the majority of bots), it might make sense to explicitly look for a counter to it, though. But hey, I'm not complaining :P

Edited 6/22/2014 19:39:27

"It's a great day for me -- first time in warlight on #1, and at the same time Poker #1 :) "
congratz :)
Its a pity that poker thematics not interesting to me..

http://theaigames.com/competitions/warlight-ai-challenge/games/53a7abda4b5ab20b4222821d

http://theaigames.com/competitions/warlight-ai-challenge/games/53a79d5a4b5ab20b42228103

Gadzbot to the rescue, dominated!

GreenTea too: http://theaigames.com/competitions/warlight-ai-challenge/games/53a7b25e4b5ab20b422282a9

Think that's why I'm so weak against the Europe strategy. Seemed a weak strategy and didn't give it much thought, but it dominates when you're not prepared to counter it.

Edited 6/23/2014 05:16:21

http://theaigames.com/competitions/warlight-ai-challenge/games/53a9e1e54b5ab20b4222a7f4

What can I say, I laughed :)

There are better ways to deal with Europe :-)

I agree:

http://theaigames.com/competitions/warlight-ai-challenge/games/53a953454b5ab20b42229f3f

Although in this particular game my bot decided to check North America first, then keep on keepin' on.

http://theaigames.com/competitions/warlight-ai-challenge/games/53a953454b5ab20b42229f3f
haha. here is hardcode nature at the glance.

http://theaigames.com/competitions/warlight-ai-challenge/games/53a953454b5ab20b42229f3f
haha. here is hardcode nature at the glance.

Nope -- my bot was trying to find the other player given I only knew 1 region he had and not 2 picks in obvious locations (best super regions).

The algorithm sought out the regions in the engine's pickable starting regions given that I expected to have more income than the other bot on turn 3 (finishing SA wasn't a priority).

Although, there was a flaw there in that it didn't check Africa first, but it still worked out :)

@Trogtog: I don't think GreenTea meant your bot.

Anyhow, here you got crushed by a really hardcoded bot:
http://theaigames.com/competitions/heads-up-omaha/games/53ab7eeb4b5ab20b4222c400
He is called OneLiner since he only consists of one line of code (apart from the engine stuff). I never played poker in my life, don't know the poker rules and have little intention in learning them.