Learning trajectories via encoding-decoding


I’m implementing a network that is supposed to learn trajectories. Basically, I created a dataset in which the training pairs are as such:

  • Input Images : a 15x15 matrix in which every value equal to one is walkable, values equal to 0.5 are beginning and starting points and values equal to 0 are walls, hence unwalkable.

  • Output/Label : 15x15 matrix, using the same features as the input images with the addition that a solution path (obtained using A*) was used meanwhile. Thus, instead of only two values at 0.5, you have the whole path equal to 0.5

For each case, I’m reshaping the matrix to a (1,225) shaped vector, sending it to the hidden layer and then to the output layer that finally compares it to the label.
So far, I have poor results. Anyone has any idea why it doesn’t work ?

Thanks !