This application trains a neural network (TommyBot) to control the vehicles in Grand Theft Auto Vice City. It was developed using Ubuntu 16.04, Python 3.6, Keras, & TensorFlow and trained on a GTX 1080 ti.
Using the Keras pre-trained VGG library, the model achieved a training accuracy of 80% and a validation accuracy of 75% after 6 epochs (10 minutes) using 11K images. Next steps are to add more images and try some of the other built in Keras models. GTA could also be a good application for reinforcement learning.
The really cool thing is that you can use the code to train any game or application that you can run on your computer. It could potentially be used as a replacement for OpenAI Universe which I believe has been abandoned.
A sincere THANK YOU to Jeremy and Rachel for this AWESOME course. A year ago, I couldn’t even spell VGG.
The original version was developed by Sentdex based on GTA V with Windows and InceptionNet.
UPDATE: 40K images and a reversing routine if the car gets stuck. Almost 9 minutes before it explodes!