Hey team
Thank you so much for releasing the code for your paper. It is very useful. I have facing one problem with the code.
I am training my code using a series of tasks. The code works fine for initial tasks but once in a while it would crash with the following error:
v = quadprog.solve_qp(P, q, G, h)[0]
File "quadprog/quadprog.pyx", line 104, in quadprog.solve_qp
ValueError: matrix G is not positive definite
I could think of one reason why this could happen - Lets say that the gradient corresponding to the current task is in a direction completely opposite to that of the gradient corresponding to the ith previous task. In this case, when we take the projection, of the current gradient with respect to the episodic gradient, the projection turns out to be a non positive-definite matrix (in the extreme case, the projection turns out to be exactly a zero vector).
Do you know of any other reasons why this might happen or if I could get around this problem by tweaking some hyperparams.