For Backers Only: AI building stuff

It appears to be a more generic issue:

The AI has currently no way of detecting dead ends / bottlenecks at all. It will just continue pushing forwards even if there is no way to expand any further.

This is partly a bug in the strategic pathing component, not detecting dead ends at all.

The other part of the bug is in the build template system:
It lacks rules which add penalty for building mobile units (and the corresponding factory types) if there are already too many idle units of that specific type in the area and/or the planet.
Latter one is equal to: "There are idle units of that type on the planet which could get to this location with a specific delay." This also is valid for air units, e.g. the AI will continue to push out hundreds of T2 bombers when left alone on a planet, even though it has no way of putting them to good use, ever.

And the final part of the bug:
Not properly demangling unit streams traveling in opposite directions and/or not calculating stream capacities of the graph built on top of the nav mesh.

True... however you read the very first lines of the paragraph and it looks like there's a valid complaint there -

- which you reiterated in another thread.

At about the 4:20 mark of this video, the AI starts to micro its tanks against the inferno, but then it overrides its own micro and goes after the MEX, getting the tanks killed.

@sorian At some point could you make a post about the technical stuff behind your world dominating AI? Possibly explain what inputs the AI/AI's neural net takes, how many layers the net has, and what the output is. Also how do you train its net? Do you use the backpropagation method or some super fancy method that involves more fancy math? Would be pretty interesting.

If this is currently living somewhere and I am unaware of it let me know.

PS: was listen to a podcast and I heard your name and I was like, I know that guy. (Idle Thumbs Podcast, gdc episode around #65)

I thought we were trying to get the AI to world domination stage, it wouldnt help if we gave everything away!

Lol Yea, are you a fellow thumb? It was the gdc epesode around #65 where they were interviewing everyone. In the interview for the QWOP guy he talked about how cool your supcom 2 AI was.

Might of just been gunpoint guy then. I don't rember.

I just thought of something.
It'd be pretty nifty if there was a sort of "mark territory" feature for friendly (particularly sub-commander) AI.
As it currently is, whenever I move to colonize another planet, the first thing I do is do a big area build on the entire planet with a couple fabbers, then lose like, 50% of the metal spots to an AI with 5 or 6 AI controlled fabbers with specific microed instructions.

So maybe the friendly AI shouldn't build where you're going to build?

It's a little irritating :/

So I was watching a replay out of boredom, some guy vs. the AI.
And the AI did this.

It took the commander 51 minutes to get out of that (only because the human player destroyed that mex)
The player built 4 factories and comm-rushed, and won.
WHO WROTE THIS AI?!

Would it be possible for the AI to detect when its com is stuck and have it reclaim whatever rendered the com stuck?

I finally managed to dig up Sorian's old GDC presentation:
http://www.gamasutra.com/view/news/178119/Video_Three_unorthodox_uses_for_video_game_AI.php

Go to neural networks in Supreme Commander 2.

that was a fascinating video.

@sorian how much time to you spend training the neural networks right now? how much to you expect to spend on the "done" AI? How much of a diminishing returns curve do you see while training a neural net? does there ever come to be a point where further training becomes counter productive (other than the time spent on it)?

As I understand it you need to retrain the Neural nets when new units are added to the game, however stat changes to current units don't need re-training?

Also, will there be a way for the community to train custom version of your AI for specific mods that add new things into the game (e.g. Statera)?

As I understand every little change you make requires retaining.

As I understand it the training is basically running the server without any client side bits, so until we get the server we won't be able to train modded AIs.