Bootstrapping Deep Q-Learning with the NEF

[Seanny123]

Somewhat related to #71, it might be interesting to hard-code an approximate solution to a Deep Reinforcement Learning problem, something like the Mountain Car problem.

In the aforementioned problem, two approximate solutions could be:

1. Build a network with integrators to keep track of how long the car has been going in each direction and use that to map the problem space.
2. Build a network with integrators to solve the problem and then see how much Deep Learning improves the solution reward.