A long time ago I used to hang around this forum with the user name redtea.  Red tea is an actual beverage that is delicious with milk and sugar.  It is not a political persuasion.

A few rednecks ran me off though because red is a red flag to them.  Kinda funny and stupid story at the same time.

Anyway you can evolve neural networks especially if you choose rather unusual activation functions that particularity suit evolution such as signed square y=-x*x, x<0

y=x*x, x>=0.   I don't think you would ever use those with back propagation:

https://groups.google.com/forum/#!topic/artificial-general-intelligence/4aKEE0gGGoA

I also have a kind of associative memory that might be interesting for character behavior,

https://github.com/S6Regen/Associative-Memory-and-Self-Organizing-Maps-Experiments

Maybe AMScalar is the one to use.

Game AI rarely uses neural networks and almost never uses genetic algorithms, and we don't really deal with general intelligence either. So, if you think you have something new and relevant, then it would be great if you could explain a bit about how you think it would work?

Game developers use to take the lead in computer science.  What happened? I presume actually work is being done on speech recognition and generation for games and for automatically animating characters.  Didn't one soccer game use AI to make the characters faces more realistic?  Two minute papers has a lot of videos that might inspire you:

https://www.youtube.com/user/keeroyz/videos?disable_polymer=1

Many of those videos are relevant to games.

There are all sorts of fancy tools at the hardware store.... but before I buy them I ask myself "what am I trying to do?" If you are trying to make game agents that can be authorially controlled by designers to do very specific things, then a hand-wavey learning system isn't going to cut it. "Let's just run this NN trainer and see if the NPC behavior just happens to match what I was thinking of when I was laying in bed last night." No... screw that. Make the character act the way you specifically want it to.

Also, remember that we aren't trying to make The Best NPCs in our games... we are trying to make The Most Fun NPCs. What's your fitness function for "fun" in your learning model?

Deep learning will impact on your sector whether you like it or not.  For example learning more natural movements for characters by imitation learning of human movement from pose estimation.  That would use 2 deep neural networks, one to do the pose estimation and another to learn sequences of movement. 

https://youtu.be/m3KG_Z0P_nU

If you prefer behavior trees and all that then up to you.

Does any of that make a game more fun? You seem to be missing the fact that Game Development is considered a part of the entertainment industry. Why throw a neural network at my soccer players face when I can just mocap/tween? Will it really matter that much to the player??? 

You seemed to be stoked about this tech, that's great, but it don't think it has the far reaching implications that you think it does. 

Sometimes developers also try to make their own lives easier - perhaps they can't get access to motion capture equipment.  If NNs are easier and more well-understood for your team than mocap, by all means use them.

44 minutes ago, Sean O'Connor said:

For example learning more natural movements for characters by imitation learning of human movement from pose estimation.

Neural networks makes sense in this case. Using computers to run millions of simulations to find good looking animations is useful; art works like that it is a process of refinement.

It's a fully automatic shotgun like approach to a blunt problem. It works very well.

Design is a precision problem. You need control and accuracy, random unexpected things isn't a unique trait it is a bug. Some bugs are good, but most are bad.

You need a sniper not a shotgun for these problems.

Neural networks will have a place in games. Mostly in crowd simulation, environment simulation and animation, where less precise actions is needed. For AI that has to respond in a controllable manor, like enemy actions, Neural networks just look like a bad idea.

It is a matter of estimation vs determining the result.

4 hours ago, Sean O'Connor said:

For example learning more natural movements for characters by imitation learning of human movement from pose estimation. 

People in games are already using machine learning for tasks like this, where it is appropriate.

For most other tasks in games, it is not appropriate.

It's important to be able to tell the difference between knowing the right tool for the job - including new tools - and chasing the latest hyped thing regardless of how applicable it is.

Put another way, there's more than one type of "AI for games". It isn't monolithic. Therefore, there could also be different "best tools" for the job for each of those. That said, when people talk about "game AI", they are usually talking about things like behavior selection and what surrounds that (e.g. knowledge representation models). For that, NNs and GAs are very rarely the best choice.