The sorian PA AI thread

It is not guaranteed that there exists a strong default choice. If you always go army first, your opponent may get to many resources.

The odds of one choice being better than the other are vastly higher than there being a perfect 50/50 split. The times when you don't shoot enemy units which can shoot back are pretty rare in PA. The AI isn't smart enough to discern such situations and is therefore better shooting the enemy troops.

When the AI attacks a base it loses battles it should win on a regular basis due to being distracted by non-priority targets. During the early game the AI misses critical fabber kills due to being distracted by resources.

Killing resources is what you do when either:

a) That group can outrun the enemy and paths by as many resources as possible

b) You've lost the battle and killing a MEX is worth more than killing one more Bolo

It's very situational. Killing troops isn't, it's a core part of the game and almost always a safe choice. Every enemy unit you fail to kill leads to a snowball effect in battle.

So from what your saying, as a starting point at least the AI should have a targeting priority as follows (for a land platoon)

Fabber > Combat Unit > Mex > Structures

Is this not something you can implement yourself with Queller?

No, AI targeting priorities are hard coded AFAIK. Whether fabbers should be ahead of combat units I'm not sure, there's numerous ways that could go wrong too, but since they die in a shot or two it's less of an issue.

Despite the fact that you have a point here:
The introductory statement is a fallacy. Say you have a 80:20 split, you might say the 80% is your default, but when those 20% false predictions make you lose all games, you might reconsider your default.

No, I feel pretty confident that the 50/50 split is less likely than every other combination of results possible, such as 49/51, 51/49, 48/52, 52/48, and so on.

Yes, but you stated that the only condition for a good or strong default to not exist is that the split is perfectly 50:50. And that is not true.

There are only three possible outcomes:

1. Shooting resources before troops results in more wins
2. Shooting troops before resources results in more wins
3. There is a perfect 50/50 split between them in terms of wins

The only measure of success is wins. Where all other factors are identical and you change this default targeting, whichever results in more wins is the superior default.

The AI doesn't always seem to follow its own placement rules. You'll see things like this:



Even though every build order for a turret always has the following in its placement rules:

Code:

{
                  "unit_type_string":"Structure & SurfaceDefense",
                  "alliance":"Ally",
                  "compare_type":"<",
                  "range":75,
                  "count":1
               }

I've seen the AI on Pacific suddenly build a vehicle factory late in the day on the starting island. Is it incorrectly setting CanDeployLandFromBase to true because a Dox from its base attacked an enemy land target?

Still seems to be a problem. Here's the AI on a user map. It produces two fabbers which do nothing but assist nearby projects. For the entire game red will never use those fabbers for anything other than assisting. It won't produce a factory (presumably because it thinks it can't afford the drain), yet nor will the Commander go and get any additional metal either.

Once the AI has put fighters in a bomber squad it doesn't seem to use them for anything else. I watched as the AI ignored two bombers on the outskirts of its base bombing MEX. It had the fighters to shoot them down, but because they were in a bomber squad it wouldn't use them.

It needs to be more flexible in its air behaviour. Fighter squads should be joined to bombers for missions, then detached once they get home.

The AI should treat Combat Fabbers as fabber units, not combat units, and it should use them for reclaiming.

You sometimes see this kind of behaviour with land fabbers on Meso. They just won't go and claim the nearby two metal expansion. They eventually get themselves over there, but it's very noticeable compared to how they behave on other maps.

The AI has a nasty habit of pausing its bombers at the end of a run if they kill their target. It's getting them killed a lot, especially by Dox. Also it's far too slow to turn its tanks around when fighting superior numbers of Dox.

There's a bug in the AI test AllMetalSpotsFull. At the start of a game it should be false, but it appears that it's never false. I also tested to see if the value was reversed by accident, but it's not true at the beginning of a game either. Any test relying on this check will fail.

It's possible it gets set to true correctly when the metal spots are full, but I didn't test that far.

The last 8 posts consist of Quitch talking to himself, fortunately he finds himself as very good company.

It's just bug reporting though. In fact, I caught masterdigital looking at it right now!
(I heard somewhere we were supposed to stalk more)

When one requires intelligent conversation one finds it where one must.

And where there is none to be had, one makes do with what is available

After a while of building on a gas giant the AI will start building its Jigs in long lines.