jim: was just thinking yesterday, will there be any some sort of live streaming event for the finals? something like analysing a few games leading up to the winners match up, meanwhile talking about the bots strategy, how this differs from regular warlight scenario, maybe chat a little with some of the coders about what their problems were developing, and with the orgas about how the compo went and what's coming up in the future.

i think it could be interesting for normal warlight players to check, could probably get fizzer to announce it on the site during the transmission to get more viewers, it would certainly be more fun for the guys competing and waiting to know about the results. and it would also be cool just for having a record of these things while they are still fresh in ours minds.

i wouldn't mind hosting it, i've hosted a few warlight live tournaments before. would just need to figure out dates so that i'm available on a good internet connection, poke a few people to join in during the event for commentary / interviews and coordinate with you orgas to somehow to avoid spoilers.

maybe chat a little with some of the coders about what their problems were developing, and with the orgas about how the compo went and what's coming up in the future.

That could be fun :)

http://theaigames.com/competitions/warlight-ai-challenge/games/5378db8d4b5ab23b41685041

Enjoy crazy game %)

what's your guys threshold for a safe hammering attack?

mine is currently 1.15 the estimated armies of the enemy region.
it used to be 1.1 but i had lots of games where i kept hitting them with just 1 army more then the defense and not do much damage, so it would just end up levelling the army count on both sides. i also tried 1.2 but after looking at a few games i felt i was giving the opponent much too leverage to expand into north america when having them pinned on south america.

not sure if the optimum value can be estimated using standard warlight attack simulation. has anyone tried it?

or does anyone has any ideas on how to calculate the best value? i thought of doing a batch test of a huge quantity of games just changing this parameter to test statistically if there is a significant advantage to having it higher or lower but i just been too lazy to actually code that. guess i still have time to try that during this week if no one speaks up.

Edited 5/18/2014 17:53:08

I made simulation of repeated attacks when I always deploy all my income to attack, and enemy always deploy all income to defence. And if after this 'virtual' repeated attacks I see that enemy is broken - then it is some sence to attack. So no magic numbers in my case.

the number of army loss has a random factor, how do you know if your virtual simulation is reliable or not? you could have succeeded out of high luck (like getting a 2vs2 victory).

Considering average values of losts. Yes, sometimes luck is not on my side :(

Edited 5/18/2014 18:36:40

Hello

What do you mean ps? With 100% luck there is never an attack that will give you guaranteed a good outcome.

If I only have one territory with which I can attack an opponent spot then my formula is:
(0.6*attackingArmies) >= (0.7 * (opponentArmies + guessedOpponentDeploymentOnTerritory)) (same formula as I use in my everyday warlight play) With more than one territory you have to write a couple lines of code. This is one of the parts where an AI can beat humans since I doubt that any human player calculates the exact expected outcome when attacking from multiple spots.

This formula is correct but the problem is that sometimes you want to also attack an opponent knowing you will lose more armies. This is the case when you can have a follow up attack next turn that will then kill more opponent armies. The problem is that such situations are quite difficult to safely identify although it's an important task of a good bot. This helps to break what I call fake stalemates. Here my formula / algorithm however is:

Attack if the ratio of ownArmies:opponentArmies before the attack is lower than the ratio of (ownArmiesAfterAttack+ownDeployment : opponentArmiesAfterAttack + opponentDeployment).

Edited 5/18/2014 19:32:37

Norman: well, yes, there is always some luck, i was just commenting specifically on greentea's approach. considering for a moment that luck was 0%, there is still a certain ratio below which is not worth attacking because you'll never manage to break the enemy even if you do have more army stack then they do and both of you lose the expected number of armies. so it's debatable if it's worth to "try your luck" with it when it's something else, but it's still important to make sure you are always above that threshold, else you might just be wasting armies ramming a stack you'll never break. i just tried random values on my bot until it felt right and ended up with 1.15. apparently you're doing the real math which comes from 0.7 / 0.6 = 1.16666667

so i guess i should be using 1.167 instead of 1.15

not sure if greentea's approach isn't better. i mean, if you're attacking with for example a stack of 20 with 22 armies and you end up with 8 9 (fake numbers, i didn't try it) it wasn't much worth attacking. you just lost a 2 armies advantage which you'll never get back. you could have used it for something else like expansion.

EDIT- just read your post again Norman, and you are considering the ratio, have to think how to implement something similar in mine.

Edited 5/18/2014 21:04:09

http://theaigames.com/competitions/warlight-ai-challenge/games/53792f894b5ab25857384a21

turn 18, 52 armies unable to take 29. 28 defender losts, 24 attacker losts!..

Update. I checked, only ~ 77% average chance to capture, so not very special.

Edited 5/18/2014 22:32:38

Hello

My bot is a 0% luck bot btw case you haven't noticed ;) I always expand with 3v2 and have no luck buffer in my calculations. I have to admit that I don't even know exactly know how to calculate the luck factor anymore without reading stuff like the Probability theory article on wikipedia. In WarLight I always use the analysis tool. I also don't know how the rounding luck in WarLight is calculated. I guess in theaigames.com no rounding luck is involved so the 100% luck feels less painful than 75% luck weighted random in WarLight.

However someone who's more into maths should be able to give a formula of the kind "if the Luck factor is A% and you attack with B armies a stack of C armies then you have D% chance of getting a good outcome (= killing more opponent armies / breaking opponent bonus)". Then you can adjust your strategy to something like that you accept a 20% chance of the opponent breaking you with a full force attack. I guess the 100% luck in theaigames.com means the same as when you play classic risk where you throw dices.

http://theaigames.com/competitions/warlight-ai-challenge/games/5379a4274b5ab25857385396

turn 17. 16 unable to take 7.. How I hate this luck...

Wow, lots of new posts since I last checked...

Anyway, turns out I'm not going to have time to finish my bot. I'll post some of the design ideas:

- The Bot tries to capture first Australia, then South America, then Africa and then the rest of the world. Apart from Australia and SA, other bonuses are too easy to spoil.

- Picks are always: all in Australia, all in SA, all in Afr. That gives 6 picks.

- Capturing Australia is most important. Even if you lose all territories in SA&Afr, you can march from Australia to that area in a few turns. This I hard-coded into the AI. It turns out that programming the opponent AI prepare for a big army attacking from Asia is difficult. Therefore Australia is better than SA.

- Deployments are determined by priorities. Australian territories as well as SA territories receive high priority. Territories with an enemy bordering receive a quite high priority. Territories with only friendly neighbors receive a very low priority.

- Attacks only happen if 7:6 advantage or higher is to be expected, even after 5 extra deployed troops by enemy.

- Troops stuck in areas with only friendly neighbors will be transfered to better places. Partially I hard-coded this, like from other SA regions you always go to Brazil, from Brazil to North Africa and so on. If no hard-coding present, then a breadth-first search will find the nearest enemy territory and transfer to that direction.

- Deployments to areas that already have big troop stacks are preferred.

- Estimate opponent income at 5. If it's more than 10, then the game is lost anyway. Income estimation by other means too difficult.

- Superregion importance hierarchy: Aus SA Afr Eur NA Asia

These were some of the most important thoughts I had when designing AI strategy.

Strategy hardcoded for map can give results, but for my oppinion it will be not true AI. I propose to raise level of abstraction and invent AI which will perform good in any maps. At least take a try. Maby not for this finals, but as ambitious task for future. Or, as variant, to make some generic core, and all map specific stuff separate in plugable module. So that bot can play without it, but, when map module is enabled - power of game on specific map will increase.

Up to now I had absolutely no hard coded strategy in my bot. But today I added a little, do deal with a few annoying scenarios like people rushing for Europe.

Is that lame?
Yes.

I want to win this though, like everyone else, so I'll swallow my pride and just toggle it off again after the finals :)

By the way, good luck everyone. I'm on holidays now and will be throughout the finals, enjoying the sunshine in a beautiful place in France and trying not to think about functions and strategies :D

Is that lame?
Yes.

I want to win this though, like everyone else, so I'll swallow my pride and just toggle it off again after the finals :)

No judgment here. I've started hard-coding against certain annoyances as well... I've even named a couple functions after the offending players :D

Trogatog:
If memory serves, you're not allowed to have any opponent-specific behaviour. Please double check the rules and make sure you're not breaking them!

Trogatog:
If memory serves, you're not allowed to have any opponent-specific behaviour. Please double check the rules and make sure you're not breaking them!

This is impossible to know. The only information you're ever given about your opponent is that they might be "player1" or "player2"

Annoyances I code are behaviors that I see that screw me. They're not specific to a certain bot, but I see them more in certain bots that I've played a couple hundred times.

Edited 5/21/2014 02:56:18

Looks like first 6 places of ladder are almost equal in game power. So if nobody untill 25 of May will not invent some 'super weapon', then it will be very tough competition.

Edited 5/21/2014 11:43:13

If Herz was better programmed this bot would have by far the best chances ;) At present he doesen't even know how to move stacks to the border though.

Edited 5/21/2014 13:09:57