GDC 2018 Flash Forward: 'Race for the Galaxy': A Neural Network in Production

what is it that makes some games super replayable over time I think it's all about offering your players a consistent challenge that's gonna grow with them as their skill level grows and one way you can do that is by offering a really competitive AI race for the galaxy is a digital game based on the original tabletop board game and it integrates a neural network driven AI originally developed by Calvin Jones and what's really special about this AI is that it doesn't it doesn't use any human generated training data it generates all of its trading data by itself which frees it from human bias which is great because human generated strategies can be fallible now not all games are gonna be good candidates for a neural network like this you're gonna want a game with high stochasticity meaning lots of random space for a neural network to explore games with a fixed number of inputs you're solving a very specialized problem with your neural network and games were canonical configurations of pieces and space it is not critical to determining who's gonna win your game it I'll get into some of the math on that at the upcoming GDC talk race for the

galaxy and neural network in production so you should come check it out if you're interested in learning a little bit more about temporal difference learning for a board game AI