You are talking about prod and this is exactly the behavior I expected ;) It makes 0 sense to play a blockade if you have no model about what is going on next turn (and I randomly guess for prod also not what is going on this turn).

Edited 4/13/2016 18:24:37

I like the idea of an ai being able to play on the ladders. I also like the concept of autopilot mode.

But I think there is an important distinction to make between an ai playing on the ladders from an account that is clearly labelled as an ai, with it's stats and performance clearly visible, and a human being able to pay a subscription to enable the ai to play for them and effect their own statistics.

For me that has the potential to demean the achievements of players and is a little short sighted. It is probably true that right now an ai could be designed that would beat the best players here. Obviously no one is going to invest the vast sums of money and time in developing an alphalight. But who can say where ai will be in 10 or 20 years time? What once would have taken a supercomputer and a team of programmers could well be within the reach of a decent computer, and a 12 year old kid with too much spare time on their hands.

With regards to autopilot in general I feel there are some very specific situations where it is a great tool at doing the hard labour in a game that is already won. Since the types of games where that is of most use are generally involving very large maps, possibly local deployment, or other less common variations, I think it makes more sense for it to be used for unranked games. This takes away the controversy of potential abuse, and lets people make use of the tool as a means to remove tedium. At the very least there should be a clear warning on ranked games if the option of allowing autopilot is allowed, and probably some way to show at any time during the game if the autopilot is turned on.

https://www.warlight.net/MultiPlayer?GameID=10938314

^This is how Wunderwaffe plays and beats an n00b player. All orders are made by the AI. The 1300-1400 rating estimate is probably accurate. The AI's strength really is in picking (always 6 picks helps a lot...) Thus it gets initiative that may be big enough to win, even with poor AI play.

I already changed the expansion a bit but it still needs a major update. Together with me making the fog removing component stronger (options are limited due to bot having to be stateless) this should be enough for some nice play.

Edited 4/13/2016 19:25:16

Norman, great work on Wunderwaffe, the AI seems already good enough to beat low tier players and finish off won games. The only problem is that it still tends to crash quite often when there's cards it doesn't understand in a game. Is it possible to make it simply ignore those cards and randomly discard one if needed?


======

For the use of the autopilot for competitive games, I think it should be probably limited in some way. I like the proposal of beren in the other thread to keep only Prod 1 for ranked games and keep the other ones for practice games only.

The only problem is that it still tends to crash quite often when there's cards it doesn't understand in a game. Is it possible to make it simply ignore those cards and randomly discard one if needed?


I haven't looked into those crashes yet. However I have tested my version on github with all kinds of cards and it never crashed. I only saw a potential problem with teammates playing a card while the AI was calculating.

Norman, great work on Wunderwaffe, the AI seems already good enough to beat low tier players and finish off won games.
I have higher aims:
https://www.warlight.net/MultiPlayer?GameID=10825463

Edited 4/13/2016 20:35:10

Hey Norman, I played around with Wunderwaffe, there should be a video still up on my twitch channel ( http://twitch.tv/wargaminglive ). Anyways I noticed something that Wunderwaffe does in regards of the current 1v1 ladder template. It seems to always pick bonuses of 3s and rarely ever combos. It does this on 2v2 ladder as well (though I'm aware it's not really thinking about the dynamics of a 2v2 game). Wunderwaffe will pick pretty much straight up solo 4s.

I was wondering at how the logic for picking is going on Wunderwaffe? Generally you should try to make the AI pick in a way it will have the most income within 2 turns.

If you could get the AI to start predicting where the opponent's picks are then that would be a huge step in the right direction.

Im not sure the point of AI is to be as good as humans (and why not better in the future ; pls don't assume it can never happen, humans operate with some logic, it is possible to record the logic on any possible scenario in AI's - unless it delays it too much, probably the only limit) ?

If you could get the AI to start predicting where the opponent's picks are then that would be a huge step in the right direction.

I can do almost nothing here. Let's say my bot gets his picks 2,3 and 4. It's very easy to determine that an opponent is in his pick 1 then... however I have no direct way to store this information since I'm not allowed to pass any state between the turns. Apart from that my bot used to be pretty good to determine where the opponent bonuses are (after they show him their income), however not under the "no state" prerequisite.

I was wondering at how the logic for picking is going on Wunderwaffe? Generally you should try to make the AI pick in a way it will have the most income within 2 turns.

Yeah, this is possible however it's not "low hanging fruit". The problem is that the expansion algorithm would have to be capable to look those 2 steps into the future. When a comparisson of multiple possible bonus combinations is involved you have to be quite carefully in regards to the runtime. My bot has some huge problems priorizing into which and how many bonuses to expand and I'd rather fix it with a "semi OK" solution which makes the bot move "kinda" smart instead of it exhaustively searching for a perfect solution. The current solution for picks is very simple: My bot picks the bonuses from best to worst. There are many factors influencing a bonus value with the main one being something close to the standard "income / neutrals" ratio. Imo this works kinda OK. My next planned step is indeed to fix the expansion priorization however my aim is to transform it from rubbish to "kinda OK". The reason you see those overkill stacks of my bot to take a single easy bonus is that my bot never expands into more than one bonus at a time (as long as he can't complete the bonus this turn). This should be quite easy to fix.