You no longer need to train a new architecture just to see if it learns. Turns out, you can estimate its success before the first training step.

By letting the agents modify their own memories, their robustness to noise becomes an intrisic characteristic.