This is in fact a kind of semi-supervised reinforcement learning model for the MIT DeepTraffic excercise.
The user is able to control the car with the arrow keys.
Additional keys are s to activate the supervised mode and l to activate the interactive learning process.
The car can learn in the supervised as well as in the reinforcement learning mode.
When supervision is active the car is rewarded for following the actions of the user. On the other hand, when supervision is turned off, the Q-learning approach is followed as implemented in the baseline DeepTraffic. One idea is to take over whenever the car gets stuck in a difficult situation. Another idea is that it will be able to learn faster in the beginning when it is totally untrained.
Download the semi_supervised_model.js file and load it into the MIT DeepTraffic website.
It is possible to set the simulation to fast mode.
Tested with version 2.0.