Evolution Strategies and Genetic Algorithms Policy in DRL

[bturkoglu]

thank you so much for nice job. I want to implement one of the algorithms without gradient in this project and compare the results with the algorithms in this project such as actorcritic, dqn ,reinforce.

I have a code that works in Pytorch https://towardsdatascience.com/reinforcement-learning-without-gradients-evolving-agents-using-genetic-algorithms-8685817d84f.

Deep Neuroevolution: Genetic Algorithms are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning https://arxiv.org/pdf/1712.06567.pdf

Evolution Strategies as a Scalable Alternative to Reinforcement Learning https://arxiv.org/pdf/1703.03864.pdf

How can I do the implementation?

[kengz]

Hi @bahaTRKGLU I have looked at evolutionary methods a little, but the main challenge is the API. The lab is able to implement all the algorithms within a shared framework because they have a common API - the control loop here.

This means for any implementation, it needs to obey the agent API. A simple example that's non-gradient is the random agent. If you're able to make evolutionary method conform to this API then u can directly plug and play it in the lab.

Interested to see if you have a design in mind!

[bturkoglu]

I basically want to implement the code shared by uber and compare the results with the algorithms in SLM-lab. But I'm a rookie in this regard and I couldn't. Can you add this algorithm to SLM-lab at a convenient time?

[kengz]

Unfortunately there's no plan to do so and we the authors are quite occupied, but I'll mark this as help wanted for anyone who wishes to take it on.