[3.1+] NullBot: an adaptive skirmish AI

[Emdek]

OK.
I would like to see it, it would be kind of non canon campaign extension.

[NoQ]

Uh oh, it's so non-trivial ...

By the way, i made something like that before: did you check out the "Cataclysm" challenge? It's for 2.3.9, unfortunately (:

[Emdek]

Nope, I was away for more than year.

[zany]

in testing this I sometimes get 2+2=5 button flash I don't know why?

[NoQ]

in testing this I sometimes get 2+2=5 button flash I don't know why?

I sometimes get it without playing the mod. Do you have enough statistics to show that it's actually more frequent with the mod than without it? I didn't test it very much in multiplayer yet, but i already had a multiplayer game that finished correctly without any desync.

[NoQ]

Finally sorted out the issues (thanks Per) and posted the "Anchor" challenge here.

[NoQ]

v0.12 uploaded.

Changes:

• New personality: NullBot ~Generic~. This AI doesn't follow any particular research path or prefer certain weapons, which makes him as weak as similar to the original AI. Not for competitive play, just for fun. I wish i could make more personalities, but #3147.
• No longer be afraid of building AA simply because VTOLs that gave the idea are around;
• Fill the biggest group first, which allows to recover faster after losing a combat group (applies to VTOLs too);
• More research path polishing;
• Building AT and AP defenses simultaneously, also allow using multiple propulsions simultaneously;
• MR doesn't overproduce trucks anymore, makes him a bit stronger in early-game;
• More efforts to decrease unit dancing.

[NoQ]

v0.13.

Changes:

• Now goes straight to the enemy base instead of hunting oil derricks when has 50% unit count advantage (made MC bot stronger in early-game);
• Doesn't build sensor towers until long-range artillery is actually available (in fact, checks if CB turret is available).
• Full research path for the Generic AI (it no longer builds gauss cannon vipers from time to time due to random research); it also uses ECM mortars and Nexus link towers;
• MC bot now actually researches bombs;
• Every bot now researches dragon bodies and base structure upgrades;
• More strict build order procedures, fixed some little deviations, as well as some production and research delays;
• More regrouping tweaks;

[Iluvalar]

The "generic" idea in V3.1 is sadly dead at the moment. Because it's too easy to overspecialize hidden in your corner and give a single do or die assault once your timing is good. Not leaving any clue to the opponent of what is going on. And choosing your research/production ratio just perfect to kill an enemy that would try to push in early game. (The only one that could see what is going on).

It's a matter of days now before the play style in V3.1 completely break from V2.3 .

[NoQ]

I don't really understand what are you talking about. The generic AI is not designed to provide a challenge, but rather to be fun to destroy and resemble the original AI in this sense. Of course it is easy to defeat him since it doesn't specialize at all. Overspecialization is the priority of hover and flamer personalities.

P.S. I wish i could do a random AI that picks one of the personalities randomly, but i didn't find any elegant way to do it yet.

[Iluvalar]

I mean that, when the players will start to absorb the experimental datas, and will realize that the lost of the unit preview reduce their chances to counter vs bad odds. They will spend less time in the alternate line that they are maintaining in hope to have a chance to counter. They'll become faster and stronger in their own path and therefore will reach the threshold point where the opponent (despite all good wills and efforts) don't have the time to adapt anymore.

Causing those others players to give up their now useless adapt path. And this will create a domino effect in the players habits.

To take your exemple, a flamer personality will now know that if he face a good mg player, that player forgot about the sensor line to focus exclusively on machine gun. The flamer player will know that even if he switch to hover, the machine gun player will shred him into pieces. He will accept the fact that, if the opponent chose mg at start, there is no hope in any way. Therefore, he will not even care about staying on wheel. He will focus solely on flamers and will send a do or die mass attack. Leaving any borg player or rocket player helpless along with any players that would have tried to keep an adapt line.

[NoQ]

Well, you're talking some right thing, but NullBot is not good enough yet to understand that; in fact, it doesn't adapt his research yet. The only adaptation so far is AT/AP weapon mix based on enemy tank/cyborg count. So i'm not sure how to apply that.

v0.14

Changes:

• Check for CB tower instead of turret when deciding wether to build sensors;
• Groups remember their targets, don't forget them when regrouping or engaging enemy forces; adds some more regrouping capabilities.

[Iluvalar]

No I know, but you try to reproduce the game experience we have in multiplayer I guess and you want you AI to be a challenge. That's why I say you should do the opposite of the "generic" AI.

But nvw...

[effigy]

IIRC the DyDo Ai had personalities that were chosen by random. Maybe you could look at it for ideas?

[Iluvalar]

the stock AI do in 2.3.9.