REINFORCE in DI-engine?

[cpwan]

The REINFORCE algorithm is a classical policy gradient method. It has been implemented in some other RL libraries, such as RLlib and Tianshou. I wonder if we can also have it in DI-engine.

[PaParaZz1]

It is not difficult to implement naive policy gradient in DI-engine, but I want to know why you need it? We didn't add naive pg before because it shows poor performance in most environments.

[cpwan]

I am doing RL on combinatorial problems. The prior works in this direction use REINFORCE. It would be nice if we can benchmark it.

[PaParaZz1]

Got it. I will implement naive pg in next two weeks.

[PaParaZz1]

I have implemented REINFORCE in this #544.