Doubts about acer HOT 7 CLOSED

@Kaixhin, thank you for your input. I will look into the chainerrl and OpenAI baselines to get more insight about the implementation. I will just send a PR if needed.

Thanks

from acer.

The trust region function should be a translation of the one from ChainerRL, apart from the fact that the trust region involves some parameters, so if you think it should be the other way around then you should raise an issue there to let them know too.

from acer.

I think since we have already defined the kl divergence before, we do not really need the negative sign. I am not sure why there is a negative sign.

Please let me know your views. Do you think it should be the other way round?

from acer.

I'm not sure either, I just ported the code from Chainer for this part.

from acer.

Let me know the difference in your implementation and OpenAI baselines.

from acer.

I've not compared the two at all, and this is very low priority for me at the moment.

from acer.

I am closing this issue because the current implementation seems to be correct. However, I think we need to detach the z_star_p for better results. Please refer to #13 .

from acer.