Dot Wars AI studies (KAIK vs RAI vs AAI)

[hoijui]

I think its a good idea to use/test AIs with bonous, case it really is needed for advanced players. Also, if the AI plays worse because of that, something is seriously wrong with it.

If one of you tells me that your AAI BA config should replace the main one, id do it.

[Master-Athmos]

AAI using the alternative config:
http://www.youtube.com/watch?v=Eixf7XXYqUc

Well decide for yourself but imo it got worse instead of better.Apart from that it shows the usual flaws again: When the green player wipes the southern left area his units start idling. It's way worse for the pink player though which does next to nothing and keeps a big army inside its base until a Juggernaut & Krogtaar arrive to crush it...

[Caydr]

Pardon me, this isn't my area of greatest expertise. Isn't AAI a "learning" AI? If that's right, then to be fair shouldn't it theoretically be best off after playing the same opponent AI a few times? Or am I way off base?

[Neddie]

AAI basically learns to react by developing values, if I recall. With too much play all these values normalize and the learning becomes valueless. With too little play it has default behaviour which is more intelligent than anything it can be taught.

[AF]

If AAI were a commercial product its learning AI claim would have been struck down by marketing standards as misleading along with all other AIs making that claim.

AAI uses statistics that it tweaks as it plays based on what happens in the game, a who killed what rota basically. It just happens to be in more depth than the other AIs in that it categorizes.

KAI sidesteps this by using an algorithm to generate the values, rather than using a default and tweaking it over time till it settles

[1v0ry_k1ng]

just to put this out there

1) KAI was designed primarily to play EE.
2) KAI used to beat top players 1v1 at EE- without cheating
3) KAI can still beat every other AI at EE
4) KAI can beat your face at EE
5) KAI degrades every spring version

6) AAI was never designed to be competetive- it was designed to make a little base and launch AI-ish little waves. it is effectively the 'easy mode' AI for noobs who fall over themselves trying to place a mex.

that just leaves RAI, which while capable of building up forces relatively well, it is not intelligent in their useage.

AAI uses statistics that it tweaks as it plays based on what happens in the game, a who killed what rota basically. It just happens to be in more depth than the other AIs in that it categorizes.

that is learning. it records statistics and changes what it builds based on the results. I always thought that a) that was incredibly ace, hats off to sub and b) why dont the other AIs do this? I bet its not hard to implement

[Caydr]

I thought AF made AAI?

I just watched the videos the first time... I'm confused on something though, what is the purpose of giving them such a bonus? What is it intended to prove if, under normal circumstances, the AIs might behave very differently?

This post did inspire me to see how an AA 5v5 RAI match on AD goes though. I'd never thought of just letting the AI play ^^ this way I never lose at least.

[Master-Athmos]

1) KAI was designed primarily to play EE.

Are you sure about that? The wiki says it was made for BA but while looking through some topics I rather got the impression that it was made as an "allrounder". It sure got some adaptions to handle the EE hubs but also was developed towards performing good in e.g. FF...

I'm confused on something though, what is the purpose of giving them such a bonus?

It's the only way not to make them totally suck. And as you actually want to fight them yourself rather than watching them fight each other I chose that mode so you can see how they perform and what letdowns they have. When dumbing things down it doesn't change anything anyway but just makes everything more cheap to build so everything just speeds up...

I thought AF made AAI?

Submarine did AAI. AF made NTAI which is disfunctional though...

The next AI I wanted to showcase here was E323AI but it has its issues and didn't perform that well yet...

[Kloot]

The old-old versions of EE didn't have hubs. Later, when I took over KAI 0.11, one of the first things I did was add support for them (for EE 0.173 iirc) on IK's request, but it was indeed built around that game.

The wiki says it was made for BA

No it doesn't? Just that it plays BA best, which imo isn't even true.

5) KAI degrades every spring version

Thanks, your quality feedback is most appreciated

[Master-Athmos]

No it doesn't? Just that it plays BA best, which imo isn't even true.

Hmm well ok I guess "most competitive AI for BA" gives you a bit too much of a room for interpretation...

[JohannesH]

I'm confused on something though, what is the purpose of giving them such a bonus?

It's the only way not to make them totally suck. And as you actually want to fight them yourself rather than watching them fight each other I chose that mode so you can see how they perform and what letdowns they have. When dumbing things down it doesn't change anything anyway but just makes everything more cheap to build so everything just speeds up...

I still think thats stupid. If they havent been specifically designed to use a bonus, then just multiplying the number of AIs instead of an AIs resources should be better. MAYBE in a human vs CPU match its better to multiply resources too, but here its no point altering it cause then you lose basis to do any comparison to human players at same point in game.

[hoijui]

Bonouses for AIs do always make sense.
Eg, if you have only AIs of one type fighting each other.. maybe you have a small map and they would not do anything (less action) if you add 5vs5 without bonouses.
Maybe one AI just does not stand a chance agaisnt an other, and you still want to compare their play styles, so you give a lot of bonous to the bad one, so you can follow the match longer.
.. there are so many other reasons/situations where it makes sense. It is not just about How well would this AI play 1vs1 agaisnt a human wiht default bonouses on a small map, ....

[JohannesH]

I disagree... Bonuses can make sense but definitely not always.

Yeah 5v5 on a small map would be retarded, but giving 1v1ing bots (of same type) bonuses for the sake of getting more action/resources/units is also stupid - they will play the map in a different manner from what was intended, and as I said then their action cannot be compared to what a human player should do in the same situation, which makes developing them harder as far as I can see, and the game less interesting to watch (as theyre playing a stupid metal map).

[hoijui]

they will play the map in a different manner from what was intended, and as I said then their action cannot be compared to what a human player should do in the same situation

It is not just about How well would this AI play 1vs1 agaisnt a human wiht default bonouses on a small map, ....

you are interested in just this one thing, which is not the only interesting thing in general. it does nto make sense for normal human players that want to play against AIs, as they can do everythign that Master-Athmos did here too, and they actually shoudl do this, to get exactly the kind of game they want.
5vs5 on a small map is not retarted. depending on how the AIs are written, it cna be more interesting, eg if each one of them has 1 or 2 attack groups, there are 5 to 10 small attack groups with 5 AIs, while with a single AI, you would have only 2.
Logic also telsl us that.. as we have these options, and they make sense to be used by users, it is good to use them for tests, as then you will see wich AI faisl with which setup, and the devs can improve its capabilities, or fix it, if it fails hard.
your picture is just way too narrow.

[Master-Athmos]

Actually the AIs won't really play different with a bonus but just faster as the bonus just affects costs making them e.g. build additional labs sooner etc. leading to being more powerful which is what you actually want when playing against them...

their action cannot be compared to what a human player should do in the same situation

Since when was that the goal here? This is not about showcasing players how to actually play the game. It on the one hand should be a help for the devs to see some not so good areas in their AI's performance while on the other hand being an interesting source of info for players who want to have a look at how well the respective AI's perform (+ in comparison to each other)...

(as theyre playing a stupid metal map)

What? First Comet Catcher is no metal map, second it's a map with next to no terrain which AI's cannot really handle after all, third Comet has a long history of actually being a good map for competitive gameplay...

[JohannesH]

My line of thinking was, that the AIs should behave as much like a good human player as possible, not to teach humans how to play... You cant make AIs mimic top players as well when they play even test games from a different position. If you make their "normal" play as good as you can (even if it wouldnt be too good), my senses tell me that their play with bonuses should get similarly as well. If you just look how they perform with bonus, to me youre not seeing as in-depth to their actions as you could.

And ofc by metal map comment I didnt mean ccr, but the fact that they get more resources than is intended.

And if an AI is coded to have just 1 or 2 attack groups, thats a big flaw in itself (and easy to fix id guess).

[Dragon45]

This is the coolest thing anyone has done with spring, ever.

Edit: Holy crap, did I see a comm-nap ? http://www.youtube.com/watch?v=ILsQlqI4 ... re=related (left side, maybe 25 second mark?)

Edit: crap, another one! cool!