alirezamika/evostra

A fast Evolution Strategy implementation in Python

evostra's Issues

What if all rewards are the same? np.std will return 0.0 and rewards will be nan

evostra/evostra/algorithms/evolution_strategy.py

Line 58 in 3819a3c

rewards = (rewards - np.mean(rewards)) / np.std(rewards)

Add Early Stopping

Add distribution (multi-processor support)

One of the strengths of evolution strategy algorithm is its linear scalability.
Thus it would be great to have this implemented in Evostra.

Add GPU support

Since one of the greatest strengths of Evolution Strategies is the parallelism it would be a real shame not to include the use of GPUs.

https://weeraman.com/put-that-gpu-to-good-use-with-python-e5a437168c01

Great code by the way.

Improve sampler.

evostra/evostra/algorithms/evolution_strategy.py

Line 44 in 3819a3c

x.append(np.random.randn(*w.shape))

Actually, it makes sense to improve this step by the following.
Let's assume that we are solving a huge optimization problem. What is the point of generating a random gaussian at each step? We can generate a huge list of gaussian at the very beginning of the training and at each step we just sample subset of given length from the already generated list of gaussians.

Moreover, if a neural network is super huge (or our objective has a huge number of parameters) it might be costly to generate (or sample from already generated) gaussian at each step. We may do that for a small subset of weights (=parameters) which should significantly speed up convergence in the case of huge parameters space.

Mirrored sampling.

evostra/evostra/algorithms/evolution_strategy.py

Line 30 in 3819a3c

weights_try.append(w[index] + jittered)

It is a useful idea to evaluate not only +gaussian noise, but also -gaussian noise.
See https://hal.inria.fr/inria-00530202v2/document for details.