numpy_ARS

A 150-lines python code for Augmented Random Search (https://arxiv.org/abs/1803.07055) with numpy.

I am definitly impressed by the simplicity and the performance of this algorithm. I did not parallelize computations, hence it could be quite less effective than authors' version (https://github.com/modestyachts/ARS).

But this version should be easyer to handle.

Quick results:

(1 episode contains 2 * n_directions * horizon environment updates)

100 episodes with HalfCheetah_V1 [step_size=0.02, noise=0.03, n_directions=16, b=16, seed=1]:

(took ~10 min on Intel i5 core)

100 episodes with Ant_V1 [step_size=0.015, noise=0.025, n_directions=60, b=20, seed=1]:

(took ~1 hour on Intel i5 core)

Running cheetah demo (after 100 updates):