We use Laplacian representation to improve exploration for reinforcement learning agents.

We accelerate Taylor mode for practically relevant differential operators by collapsing Taylor coefficients; this can be done automatically with compute graph simplifications