Extending Evolution Strategies Starter Kit


So the blog post and research paper and starter kit got everyone very excited. As someone relatively new to ML (and not with a particularly strong maths background), I wonder if anyone has been able to use the starter kit to extend it to other gym environments (especially those that benefit from network types other than FF - for example LSTMs or CNNs.

I’ve been poking at the code and the only example provided is for Mujoco - https://github.com/openai/evolution-strategies-starter/blob/master/es_distributed/policies.py - and I see that FF is the only implemented architecture.

The other thing - I mention only as a suggestion - is there’s very little documentation on how to get this working locally (for development purposes) - I have just about managed to figure out how the scripts hang together but a few instructions that are non-EC2 centric would go a long way.



I’ve developed an Evolution Strategy module in python which you can use it with any type of network.

The code is available here: https://github.com/alirezamika/evostra

I used the module to train an agent playing the flappy bird game and I got some interesting results which you can see them here: https://github.com/alirezamika/flappybird-es

I hope this helps.