New Contest - game like Risk

[GreenTea]

[jab_bott]

I took a quick look at warlight earlier on... well that was my evening gone, 7 hours later now finally I'm replying to this post!

I think the game is totally awesome! However, I don't think it would make a good AI challenge because it's too simple in the sense of the AI/strategy. I only played the single player levels, so I didn't see all the types of cards and variations, but it seems that the basic strategy is to maximize your presence on your borders and minimize it internally, find good defensive locations with only 1 route into your territory so you can focus your troops and defend it cheaply, and good attack-points (weak spots) by taking one territory with lots of enemy borders so they have to spread their defenses out... all the complexity would be in complicated rules such as how cards behave and various bonus/power-ups, which is fine for a game that you play as a human, it makes it interesting, but to write code for it the challenge becomes more about understanding a complicated set of rules than about writing AI. Also, I think the combat would be the major part of it, whereas ants had other elements like the trade-off between explore/defense/hunting/attacking, this game would be mostly focused on only the attack/combat part.

I could be totally wrong of course, the game really is great fun to play. I just think that something with a less-complete solution would be more challenging.

[GreenTea]

All you say is right. But this is only basic strategy of the game. It works only against week or average oponents. But if you play like so with stong oponent it easy beat you, because will predict all you moves, and make good conter strikes where you not expect (because use your basic default strategy). This is mind game based on prediction. During play I always learn my opponents game style and try to adopt to it. For example if enemy play very safe and defensive - I'l better not make aggressive moves. If it play to aggressive - I can let him attack me on one direction, and make conter attack in other direction, with more damage. Its just simple exapmles. Basic idea is that there are no one imba strategy, player who better predict moves and make unexpected moves will win. I ever like idea for this type of game to let bot to save some information between games. For example file not bigger that 1 MB. So bot be able to learn and adopt to other players, and get smarter in every next game.

[jab_bott]

Mmm... yes, very interesting points.. I'll have to go and play it a lot more to understand that more advanced strategies

[george]

hi, i play warlight quite a bit, to a good standard. it is not at all a simple game to play, for all that the rules are fairly simple (in the same way, perhaps, that poker is not a simple game to play - and AIs struggle with that).
you are using a lot of (maybe implicit) probability, game theory, double-bluffs, checking opponent's game history and win% to guess playing style and likely moves, etc.
also, playing style has to adapt to the map - the big maps play differently from the small ones, barriers on the map etc change the style of moves quite a bit, etc.
would definitely make a good game for a future contest.

it is probably the case that you should only do it with 2 teams (eg 1v1 or 2v2 etc) - otherwise, as Fizzer says, it is too much about diplomacy than pure strategy/tactics. - the ladder competitions are all of this type as it is the best/purest way of determining skill.

[iouri_]

Warlight looks like a fun game to work on, fits the game criteria pretty much perfectly. Same with Risk on a new set of maps.

And would be very easy to work on (but not easy to solve) on the participant side.

Would be nice though to get rid of the random element; but I guess if that's not easily done then it's OK.

[smiley1983]

Here is a working (but incomplete) prototype of a Risk-like game if anyone wants to test it:

https://github.com/smiley1983/aichallenge/tree/war

No visualizer yet, and the only starter is for ocaml. The stats make it look like it's working, but without a visualizer (or any sort of unit testing) that's little more than a guess. I might get a chance to work on it some more soon, or it might be a while.

Currently no bonus for controlling grouped territories, and combat is deterministic (60% of attackers and 70% of defenders kill one enemy). Base unit numbers are multiplied by 1000 to reduce rounding errors while only using integers.