Hey guys, so with poker, there's a lot of talk about AI/RTA (real time assistance).

Seems like anyone could easily create an AI for warzone and cheat without ever being caught.

Sooner or later it will be easy for any player to use free or very cheap AI services as RTA.

Thoughts?

Edited 3/21/2024 18:18:22

Currently, the best AIs are Prod 2.0, Cowzow or Wunderwaffe. Wunderwaffe and Cowzow were bots written for The AI Games. I think these were 1st and 2nd place while the competition lasted. All of them are still quite easy to beat. It would be impossible to distinguish between an average player and a bot at this stage. WZ might be a bit to complex to write an AI that does well against real players. Lots of things can happen in a turn and sometimes the best thing to do could be the worst, depending on which options you evaluate.

All I can do is understand the settings and have a stab of where opponent is. Where they are and what they are doing is different among how well people play. Writing a bot to do that is really ambiguous.

All of them are still quite easy to beat.

If i remember correctly, Norman said thats cause the current AIs cant access the history of the game.
When youre running a game locally, you can expand your input with way more data than just the current standings.
Just the history on its own is likely to improve WunderWaffe's abilities, and you might be able to feed an AI much more...

Seems like anyone could easily create an AI for warzone and cheat without ever being caught.

Very bold claim...

I remember seeing him (Norman) saying something about that, but I don't see why it would be limited to only using current standing and having no history.

Edited 3/22/2024 00:14:28

(double post)

Edited 3/22/2024 00:14:45

I remember seeing him (Norman) saying something about that, but I don't see why it would be limited to only using current standing and having no history.

Because Fizzer made the bot API that way.
I have just checked: the bot API provides info only for current and previous single most recent turn.

Edited 3/22/2024 00:33:38

Well, let's just assume the AI is optimized for SEAD or regular SE.

Think of it like Poker, you have your "pre-flop" gameplay (in Warzone, this is when you pick your starting locations).

Then in Poker you have "post-flop" gameplay (in Warzone, this is turn 1 until end of game).

Like Poker, Warzone is a game of math.

On SE, if you know what the possible picks are, you could ask the AI (or a "solver") for the best possible picks. And, based on math, it can tell you which picks have the highest % of winning.

How are you guys not seeing the potential here? This is already happening in online Poker.

if it's limited to current turn, then all AIs are hopeless against humans who use history to make decisions.

https://www.warzone.com/Forum/436712-think-small-earth-solvable might be an interesting read as well

A lot of smart people have worked on poker AIs a long time to make them better than humans. Same with Chess. Computing power and statistical learning are better developed than they used to be, but I would guess it would take considerable effort to make a program that played better than a human, and without history impossible.

Guys forget the bots, this is about GTO, Game Theory Optimal.

Like the thread JK posted, SE is in fact solvable, mathematically, based on picks alone.

Like we all know that most games of SE or SEAD are decided by starting location in the majority of cases.

I d be far more impressed by someone winning using AI than just using its own brain.

A poker AI is rather easy to write as calculating all the probabilities is pretty basic maths.
Obviously you still need to select a move based on those probabilities (preferably in a non-deterministic way),
but that cant be to difficult either. And still it it took some time.

Warzone has some characteristics that make writing an AI very difficult. No full information like in chess, also different board each game. A wide variety of settings. This makes a self-learning AI difficult, although surely not impossible.

In some games AI's managed to be decent despite these characteristics (some RTS games or dota-likes), at least when given limitations on settings.
I m not up-to-date on if they are stronger than humans by now in some games, but in these games AI normally makes up for its weaknesses via beyond-human reaction time, multitasking and eternal focus - non of which matters in warzone.

I d guess the stronger AI's would emerge from some hard coded rules, which i dont see beating a strong human player consistently any time soon. And if such AI is actually successful, the creator themself must be an incredibly strong player to be able to formulate such rules (unlike any evolutionary AI, where the creators could be terrible at the game itself)

I think GiantFrog is right

I've tried to create an AI myself. https://www.warzone.com/Forum/670192-ai-capable-defeating-prod-20

It's certainly better than Cowzow, and pretty close to Prod. Wunderwaffe is leaps and bounds (shoutout Norman).

I agree with GiantFrog that creating an AI for warzone is intrinsically much harder than chess or poker, but the real limitation is what JK_3 said, that AI can't access history. That's a huge deal and a significant barrier to developing an AI.

@Fizzer, pls change that :)

Do the bots have access to both the exact orders played and the current game standing?

Current AI's (bots) are quite good.*
*with no fog

For example, I took Wunderwaffe, and changed nothing of it's gameplay mechanics, but I did change the way it does picks. I added a lot of logic that specifically applied to the MME map and 1v1 ladder settings, based on my own understanding of the game.
For instance, I gave it an idea of coverage, absolute and relative bonus strength, expansion potential of picks, safety of picks, combo possibilities (f.e. 5+3+3 bonuses), and more.

This bot is, in my humble opinion, able to beat 99% of the playerbase on pick outcome. It scored a medium place in a picking competition with all 1900+ MTL players a while back. And most importantly, I myself do not manage to predict how it picks a given board.

Reason I'm saying this is this:
It's not too difficult to put the game and thoughts of good players in numbers and logic. (I'm not even a developer)

Now, I also let this bot play no fog games (remember, it's Wunderwaffe playing still).
It won a few games against very good players (2000+ MTL). And I would say it got a winning advantage after 5 turns in maybe 10-20% of games. The main issue with current bots is that they are too predictable in gameplay, and also not great at stack management. So you can almost always outplay them from a losing position in a long grinding game.

So, the upshot here is that they're already better than people believe, and also potentially self-learning ready (my pick logic at least is very number/evaluation based and therefore perfect to be optimized). That said, the variables it would be optimizing are still the ones coming from my brain, not some variables that are objectively good to be optimized. And you know how they say... Shit in, shit out.


All that said, for top play I do believe that we are still far away from bots or AI beating people. Many arguments here have been exchanged regarding that, I don't necessarily want to add.

For casual playing though, I do believe there is quite big potential for RTA even within current restrictions, but it's different than people think. The most difficult thing for bots/AI is to find their "intention" for a given turn - what should they achieve? We can humanly take that decision ourselves and then let AI optimize our moves.
For example: I am entering my moves - deploy 12 to A and 7 to B, delay some and then attack on those 2 spots with 12 and 7. An AI could take that input and optimize a few things. It would realize my intentions (attack those 2 spots and outdelay my opponent). Based on that, it could redeploy my armies, make efficient delays (using conditional delays, using the correct amount of delays, ...) and could attack with efficient numbers (f.e. 11 and 8).
Another example: In a multi-attack game I would attack a certain path towards my opponent. An AI could take that as input, and do the nitty-gritty detail work such as the multi-attack double-attack trick, which for ideal moves you have to do several times at different locations to account for all possibilities)

All of this is especially helpful in RT games. I am thinking Small Earth 1 minute, where the delay optimization could come in super handy, even for a good player.

Edited 3/31/2024 14:13:51

It's already possible to use stock Wunderwaffe to outpick and/or outplay most players much of the time in certain games.

That it might not yet beat top tier players is rather small consolation to the rest of us, lol

My mobile phone currently stores a game that I should really get back to, but I havent because I'm stuck, lol
I'm not the best player, but I've beaten quite a few top tier and second tier clanners in RT QM and RT CW games.

https://www.warzone.com/SinglePlayer/Level?ID=1507808
Small Venus Map, Custom 1v1 Pure Skill template, created by Sapin, (props),
with Wunderwaffe as the AI, (props to Norman)
(granted, many of these wins against humans were because the humans didn't bother to notice who they were playing against...
WW vs. DW at least 3-0... LMAO)

Small Venus
Created by Sapin (all)
State: Public
AI: Wunderwaffe
Went public on 10/24/2023
Number of attempts: 2,289
Number of wins: 620
Number of likes: 48
Record holder: Lepanto31 [Warzone Member] in 5 turns on 10/27/2023
You like this!
Your attempts: 360
Your wins: 256
Your fastest win: 6 turns

That's 104 personal losses, at least 99 of which I knew who I was playing...
and probably 94 of which I knew and understood the basics of how to beat Norman's bastard. LMAO