Why You Can't Use AlphaZero for TripleA
-
I found this interesting paper about using AlphaZero to play Risk.
FULLTEXT01.pdfAlthough AlphaZero based programs can beat the World's best Chess & Go players, it does not do well in 2-player Risk, playing much worse than a human and implicitly against algorithmic AI.
Some things of note from this paper.
-
AlphaZero can't figure out important things on its own. While it could learn to play Chess & Go at super-human levels, it needed to be told specifically of the importance of conquering continents
-
AlphaZero's approach still requires studying moves and outcomes. But in game of Risk, there are many many legal moves before play goes to the opponent. AlphaZero starts with randomness and there are too many moves to start choosing effective moves
-
Random combat outcomes are a big problem for evaluation. AlphaZero can't distinguish easily between good moves and lucky moves
-
-
@rogercooper break out the quantum computer
but seriously our current AI is pretty decent i think as long as the map isnt too complicated and as long as you forgive certain weaknesses currently part of its decision tree
-
@ubernaut Yeah I think the AI is not terrible, I play a lot of Global 1940 and the AI performs well and only breaks down with specific edge cases that I think can be fixed. by the way I am interested in helping out with development for Triple A , I am still trying to get my computer setup.... I need to run gradle but making progress)
AI issues I run into for Global 1940-
Germany can't see two moves ahead, American captures Denmark on it's turn opening up the Denmark strait so Britain can easily capture a lightly defended Berlin and take all of Germany's IPCS (This is a known issue on the AI thread)
This actually might be a hard problem to fix, but it's theoretically possible to fix. -
Japan declares War on some powers and not others, IMO it's never optimal and I would force them to declare war on everyone or no-one. (this could be left along for Easy AI) but I think for Hard AI it should be all or nothing, otherwise the game just starts going into scenarios you will never see in a real game.
-
Germany declares War on US in Europe 1940 before round 3. IMO almost never optimal, but their should be a fail- safe here that says (If United states leaves ten transports undefended and germany has one sub within range and expected TUV is 50 + then yes Germany can declare War.
-
AI declaring War on the True neutrals, I would love to create an option in their settings where you can prevent the AI from declaring War on True Neutrals. This really breaks game immersion when it happens. I know the Spanish beachhead is a strategy, but I hate it when my AI teammates declare War on True neutrals. Fix would be a game option setting that says AI cannot declare War on True Neutrals.
-
In Global 1940 make Anzac actually defend Australia, in my games Anzac almost always makes a beeline for Africa. If there could be some sort of game logic to prevent them from leaving the theatre until Hong Kong, Malaya, Philippines / Shanghai are Allied held that would be great.
-
Russians send half their army to leningrad and attack Finland and let the Germans breeze past them on the way to moscow.
-
America Purchasing. There needs to be some sort of calculation pushing America towards a balance of transports/ seapower / and land forces. Sometimes you see America build 40 infantry and no transports. They need improvement when it comes to building convoys to Europe , they need to use harbors better and once they get to Europe they make all sorts of interesting decisions . Like just leaving troops in random places and not being able to consistently move troops to mainland Europe.
I think Russia/ China/ Germany play well on land as does UK Pacific.
-
Anzac never learns that need to take Dutch new guinea to get their objective bonus....
-
When I play Britain I know for a fact that Germany will never try to build more than two transports and attempt sea-lion... This isn't a huge issue but it would be great to somehow fix this.. relatively a hard issue.
-
Japan really loves going after Russia and the mongolian territories. This is an example of the AI doing well on land and bad on water. I would love to have a setting that says Japan will not declare War on Russia. It would be nice as Germany to know that your AI teammate is actually going to do their job and take islands in the pacific. Japan will get around to this but it's not always their priority.
-
Japan leaving defenseless transports around goes for other powers as well. They really need to protect these better to be effective.
-
It would be great to have a setting, that allows you to pick what round Japan will declare War on the Allies. Just to try out different strategies when you are germany.
Overall I think the AI is pretty good when it plays Germany, except for the gamebreaking Denmark can-opener I think they defend Germany very well against amphibous assault, sometimes this is because they don't prioritize a Moscow at all costs- as fast as possible strategy.
I've played hundred of hours on Triple A , so just heads up I am a huge fan of it. These issues are not criticisms just things I have noticed and areas I thought could be improved. A lot of this stuff isn't essential and not truly game breaking. I think Triple A runs very well and the game rarely crashes, I know you guys prioritize game breaking bugs and a lot of stuff that users can't see. Thanks so much to everyone who has worked on it. Hopefully I will be able to get my coding environment setup and start wrestling with the code on how to fix these things
-
-
@jack-sarles yeah defending transports has always been a weak spot for the AI, another thing AI doesn't seem to understand in global is the effect of air and naval bases.
-
@ubernaut Yes that's a greatpoint, the AI doesn't appropriately value the ability to have more options through air and navy bases.
-
@jack-sarles sometimes it will park a fleet right in a sz which adjacent to another sz with a naval base which seems pretty nonsensical it also doesn't seem to recognize a fleet which is 3 spaces away and next to a base to be "in range" for an attack
-
@ubernaut Hmm interesting, I'll try to look out for the issue of them defending against units with movement boosted by air bases / naval bases. This might be related to the defenseless transport issues. Sometimes I think the AI does a good job of defending transports and sometimes not. Specifically in the Pacific theatre, which might be an issue of Japan not understanding the philipines Fighter and naval units of both manilla and pearl harbor can move 3 spaces. they can be vulnerable to this as well as a fighter attacking the sea of japan and then landing on an Aircraft carrier one space away. This also relates to the US sometimes totally abandoning the pacific theatre and focusing solely on germany. This strategy is perfect for old versions of axis and allies but not so much for 1940....
-
@jack-sarles it's better than it used to be used to not defend them at all now does sometimes
-
@ubernaut well that's good!
-
@jack-sarles yea politics and objectives seemed to be the two biggest ones for the 40 game from what I remember.
Wonder if it would be better to script something. Player A does this Player B chooses one of x amount moves.
So many possibilities, it would probably take years to do.
-
@beelee I am in agreement. The current TripleA AI is pretty good, although it has some fixable glitches. I think the further improvements should come from the ability to create specific heuristics for the AI for specific nations and scenarios.
What isn't going to happen is the creation of neural net style AI that will play well. A&A is beyond the capability of current AI approaches to solve.
