ananyakumar / cooperative_learning Goto Github PK

For each car c, we want the k closest cars. For each of these cars we want the relative velocity (x and y component) and relative position (x and y component). So we will get an input vector of size 4k [pos_x1, pos_y1, vel_x1, vel_y1, pos_x2, pos_y2, ...].

This then gets fed into a (randomly initialized) neural net with fully connected hidden layers which has 4 outputs: mean and std-dev of acceleration in x and y direction. The final layer would be linear, except for standard deviations there's usually an exponential activation so it doesn't get negative. This is plugged into a Gaussian distribution to generate 2 values per car: acceleration x and y.

The controls for the cars should be concatenated into a 2d np array, and given to the environment. If the neural net stuff is tricky, for now let's at least do the input and output preprocessing so all we need to do is build the neural net and train it :)

Add support for obstacles into the code.

• Add an obstacle class which has a center and radius.
• For now we can have a fixed x, y coordinate for the center, and radius for the obstacle.
• In the _kill_collided_cars, check if a car collided with an obstacle, and kill the car as needed.
• While making changes, let me know if the code is poorly structured and I can fix things.