frob said:
That's something that would be a poor fit for machine learning. There isn't much to reinforce or a way to get better, it is a binary state, either they made the shape or they didn't. During training a character could perform millions of pickup and placements and never create a single ‘symbol of power’.

I meant to use ML only to detect the symbol a human player has made, so pattern recognition just like reading human written text.

frob said:
But for game mechanics, where human designers need to fine tune values, humans adjust probabilities and frequencies and other weights, those tend to be a terrible fit for ML. Stick with state machines, behavior trees, and human adjustable math formulas.

It would be interesting to know if a chess ML-AI can learn to play worse than optimal, and so adjust difficulty to a range of human players to loose on purpose, but pretty tight so causing satisfaction. Probably that's no harder problem than to learn chess at all.

We will see what the future brings. New games or new genres might deal with new limitations, problems and options in ways we can not foresee yet. At least i hope things keep changing so the medium remains interesting.

Calin said:
The game developer knows all the possible outcomes of the game.

Yes, for now.

But it is not how creativity usually works.

For example, a music composer does not construct his results from precise plannings. He sits there and plays, tries this and that, and then suddenly, with luck he discovers a melody, and he knows: “Yes! this is it! People will whistle this melody, because it's an earwig.”
But it also feels like this melody comes from the outside. Was it random coincidence to discover it? Was it sent from god? Does not matter - it came from somewhere.

Game designers (and computer programmers) don't have this option of loose inspiration and random luck. They have to be too precise. Stuff only works after nailing every tiny detail that a regular inspired mind would not even had to think about.

So, if ML can discover ‘gameplay’, like learning to use ramps or to block doors, without a need to plan and program for this in the first place, game designers suddenly get this random luck moments and inspiration from outside too, eventually.

But all this is just loose thoughts and philosophy. I have no plans to use ML for game design, and i would do exactly what frob proposes. I have no interest to change this. I even thought ML is very restricted and the current hype around it bores me, tbh.
What changed my mind was the work of a researcher that managed to make ML composing earwig melodies. I always used exactly this as example of something ML could never ever achieve, but it did.

So, never say never.
But why the hell are you so very interested on the topic? What do you expect to get from ‘advanced AI’ for the game you work on, or the games after that? What is missing from current games because of bad AI?

You can`t make a game that you don`t know what its going to be.

Why not? You certainly can, it works for all other forms of art. (I do not consider games as art, but game design is.)

It's all about inspiration. In the end you sell a verified and approved product in any case, early design documents are not that detailed. Diablo 2 never had a design document for example, and it has just 'grown', they say.

I believe in ML to be useful for content creation of many kinds, but not so much in runtime use, for now.

Making music or drawing is just like programming,

No it's very different for me.

behind every creation there is a creator.

Often his name is Mr. Random Luck.

After that it's easy to present your work as a result of brilliant planning and hard work… but it may have been lucky coincidence that seperated it from all the others.

Calin said:
Diablo 2 is imitating just like any other game.

Point is it's design was pretty loose and free, and they did not knew what would be the end result, opposed to what you claimed to be necessary in game design.
The fact most games are similar to other games is orthogonal to this. Also the fact games became products of hundrets of people instead just one would not rule out ML being useful for prototyping or giving unexpectedly good results in lucky cases.

Computer learning makes sense in at runtime in games people are just too greedy over CPU resources

No. We have now gaming GPUs that have tensor cores to accelerate ML, although we do not even have an application for those cores in games.
So feel free to use ALL of their power :D
(…ignoring the fact using tensors also means using regular shader cores, and they are meant to execute models that have been learned offline.)

People are not greedy on CPU resources either, they just try to utilize them. If you come up with great AI that needs more CPU, we can tone down some other things if necessary and worth it.

The real questions remain:
What do we aim to improve, what's the issues with current state of the art?
How do we achieve our goals? (ML, or hand written optimized solution possible?)

Before we can answer those questions, the whole discussion remains just play of thoughts as usual.