For Backers Only: AI building stuff

On the good side of AI I just had a Galactic War game that went poorly. Was playing a two moon map I'd played several times before, knocked the two AIs off the main and set about invading the secondary moon. But this was deeper into the campaign so was proving a little more difficult than usual, had two teleporter beach heads shut down and I hadn't upgraded my eco because I thought I was about to end the game. Then the AI dropped a nuke on me.

Right, **** you AI, we're gonna play it hardcore I thought. Started spreading out my forces, though the next nuke took out my T2 factories. So now I set everything I had to building an anti-nuke, then I started building an epic complex of nukes. This was gonna be sweet. Got a warning the AI had launched a nuke and I laughed, obviously it didn't know about my anti-nuke.

Apparently it did. It had launched four. GG.

It seems to be pretty decent at multi-planet games and not so good on single planets, for sure!

Not sure what game you are playing but when I watch the AI play multiplanet games it doesn't even build metal extractors on it's own planets. It's totally broken

Just played a GW game yesterday where two of my support commanders got a moon each for themselves - they built mexes all over the planet and proceeded to build 20 bombers or so, then stopped doing anything, completely ignoring orbital and T2.

If the ai has a low amount of nearby mexes it becomes borked.

Awesome!

Would be cool if AI finally learn to repair it's commander. With offline released it's possible to play really long games now and this is one of main AI weaknesses.

i have seen it repair its commander, but indeed....incredibly rare to see

So watching that lecture/presentation on neural networks oyu made in GDC tells me that basically as soon as I or @liquius get a better PC and can host our own server we can totally train the Neural net to work with RCBM units, correct?

Therefor, no. Training isn't part of the shipped binary (yet).

How do threat maps for units work? If the AI spots units and then loses sight of them, how are the threat maps updated to reflect this?

Thinking that this might be the cause for some neural net issues.

Im seeing The AI loves to fly around with the commander in astreaus, and land close to enemy bases, even tho it got a big base at home planet, why take the risk?

Are you able to provide any hints on how the following lines are used by the AI?

metal_drain_check [0 1],
energy_drain_check [0 1],
metal_demand_check [0 1],
energy_demand_check [0 1],

From what I have gleamed, I think it's a modifier for
"test_type": "CanAffordPotentialDrain"
and "test_type": "CanAffordBuildDemand"

Can the AI config files make it reclaim features?

It's also sounding like I might have to disable drain/demand checks to make it build anything with no steady-state income.

Does economy_scale do anything? I couldn't see any impact from this value. Is that the old built-in economy modifier?

It's in the ai.js opening comment section.