I just wanted to share some information about my master's thesis you might find interesting.
In my master's thesis (in which I am cooperating with a fellow student at Aarhus University), I take a look at machine learning and AI through the eyes of the game-developer with the overall question "is machine learning ready to be applied by a small indie studio?" looking at not only the possibility of using machine learning, but also the feasibility.
Unfortunately, Twitch keeps deleting my videos.. So I have an old crappy YouTube video, which shows the basics of the game here: https://www.youtube.com/watch?v=83mbhlgD49Y
It is a top-down shooter, all players are AI (although ideally we would have human players as well, but alas it was not a priority) and they are limited in vision. Each tick they are allowed to choose an action based on an information vector. We had a hackathon where people submitted AI's for a tournament, after the hackathon we used these AI's for training and evaluation of our own -- machine learning based -- AI.
We have had limited success so far, maxing out at 16% wins against the best hackathon AI, and 30% against the worst, but we see a lot of interesting trends and we have a few more tweaks we are going to make.
For those who are following the field, we are using Dueling-Double-Q-Recurrent-Neural-Networks with experience-replay and a simple weight evolutionary algorithm as well as stochastic gradient descent with an ADAM optimizer. Pretty much following in the foot-steps of DeepMind and other major projects out there.
Our game source (albeit ugly) is available here: https://github.com/lukaspj/Speciality/tree/speciality , it utilizes almost stock T3D, with a C# layer on top of it.
The machine learning AI code is available here: https://github.com/lukaspj/Shooter-Learner , and is a lot prettier by comparison.
I have a draft of the thesis available at: https://1drv.ms/b/s!AtrTgygwv8wijupGwI9hoYWJrfbpfw
It's due June 15th, if you are interested and read the draft, then any comments to the content is much appreciated, but it's still very much in a draft-stage, and the TODO's are in a mix of danish and english.
If you have any questions about machine learning or other related topics, then feel free to ask them. I will be happy to answer, especially since it will help me to prepare for the defense of the thesis.
I hope it was remotely interesting! I will be happy to provide more details to anyone who wants them. I just you might think of it as an interesting story how Torque3D has helped me implement my thesis project.