TITANS PTE build 88043-pte is now live!

The issue when the screen is white is still there: (linux/nvidia)


F5 fixed it, but this is something newcomer will not know...

Log included.

@ Sorian The new AI:
In GW, in standard mode (i do not know the exact term in English), the AI almost does nothing. It can even build a factory on top of a metal extractor. Even if it has only 4 mex available in its base.

The AI in GW has been nerfed into do-nothing-state for any difficulty. I find it unplayable since the AI does nothing to fight back.

Issue with build orders:
Sometimes, only a few fabbers will respond to orders. See this exemple with building a catalist:



Sometimes, the order will not be completed:
I ordered to build 2 catalists:

The second one never built.


The full log is attached.

The workaround:
Build first catalyst => Shift click to move to a spot near second catalyst => Shift click to build second catalyst => Shift click to move to a spot near third catalyst => Shift click to build third catalyst.
And so on

This used to be possible right?

No.

Yes, it used to be possible.

Really? You sure? I don't remember that. I guess it got turned off because it could end in weird **** happening.

There is also two bugs about the catalyst.

The first one:
Player can't see what blocks the placement of catalyst!

The second one:
After deleting nearby buildings, the fabricators don't automatically clear wreckage in the placement.

I think the real placement size of catalyst is bigger than it shows.
It is just so annoying to figure out which object blocks the catalyst.
Since the placement is fixed in only five locations, how about removing their collision detection?

Still pathfinding issues while entering teleporter, Colonels have huge problems in that area

I'm a bit late to the table, but can @Sorian tell us if the AI can learn from the player and build into the AI net?
I seem to recall playing the old TotalA back in the day, and the more I played the more difficult the AI got.
It certainly did seem to be learning and growing each game. So is there a reason why it can't happen this way in PA?

Assuming it's possible, when multiplayer happens, as well as generating planets- each computer can sync ai net paths to servers. That would give a more streamlined updating system so that everyone can help improve ai together.
Since the ai is still limited to the game, it won't become skynet.

It would be interesting to have an AI difficulty that starts off at the "normal" or "hard" setting, and then learns form each player it plays against. Having a client-side version for you to play against that gets better as you get better would definitely be an interesting way to help players improve. As you get better, so does your AI opponent.

I will be very happy if the AI can really be trained by players.
But is the training always good?
Or the newbie may train the AI to play even badly?

Think about the rank of 1vs1 game.
I believe many players (perhaps over 50%) still don't know the detail of units or how to expand.
This also happens in custom game, especially in FFA.
The performance of players are significantly different.

I wonder the training of AI from every player will finally make the AI as a Sliver player rather than Platinum or Uber.

As someone who worked a lot with the Total Annihilation AI, I can assure you it was thick as **** and most certainly did not learn

The player cannot train the AI, but the neural net doesn't work the way you're describing. I recommend watching Sorian's GDC presentation on the subject.

Doesn't the AI train by fighting itself? Wouldn't it be possible (with some coding workarounds) to make it take up the same process while a player is playing against it?

The only workaround required would be enabling people to turn on training. It's currently disabled outside of Uber. You'd have to play hundreds of games though.

Yeah, that's what I mean, you'd turn on training, and have an AI that logs the results from games played against players playing against that difficulty of AI.

Training on anything other than the optimal settings would not be a good idea. Each AI personality introduces its own modifier into how it interprets the neural net results.

So you wouldn't be able to make a single AI personality that would only be used for this hypothetical adaptive AI?