Deep Learning: Train Network to "Play Game"

My idea is to train a network so that the network can perform similarly to human subjects when playing a simple "game" eg. tracking an on screen cursor with computer mouse. The goal would be to train the network on human subjects' input/output data and then have it "play the game" and achieve similar output results. Ideally the network would have similar success and failure rates as the human subjects. Is this possible with Matlab's Deep Learning Toolbox or the Statistics and Machine Learning Toolbox? Thank you in advance!

Answer 1

[Gabija Marsalkaite]

Hi Michael,

This strongly depends on how you want to implement it. Deep learning alone is most often used for tasks with labeled data (e.g. pictures of cats and dogs with corresponding labels) so if you have a labeled set of samples that might be a good idea (e.g. pictures of screen as samples and direction of movement as labels). Alternative to that is using reinforcement learning where agent learns by performing some action and then receiving a reward according to some quality of action metric (e.g. movement performed by agent is an action and reward is calculated by how close it is to the target). The plus of reinforcement learning that one does not have to create labels in advance but negative is that occasional it is difficult to define the best function describing the quality of an action.