Neural Network to Control Simulated Craft
Trained using a genetic algorithm
Individuals gain fitness scores based on their ability to collect the green waypoints and remain stable (by not spinning out of control).
We created a variety of crafts with varying control schemes, and some performed better than others. On the left is the most complex and difficult to control craft, which has two independently vectored thrusters.
In the bottom right, you can see our representation of the neural network’s behavior. For the craft shown, the inputs (leftmost nodes) are the position of the craft relative to the target, the current velocity of the craft, and the current angle of the craft. The outputs (rightmost nodes) are the angle and thrust power of the two independent thrusters. The size of the links indicates their weight, and their colors indicate the current values being added from each connection.
Future improvements
In order to improve performance, we could incorporate multiprocessing in order to train more individuals at once and allow for a realtime visualization of the training process, since training currently blocks the main thread. We also had trouble with the craft overfitting to the training waypoint locations, and we could add some dropout layers to address this. We would also like to improve the visual fidelity of the visualization tool.