search

ikostrikov / pytorch-a3c

PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from "Asynchronous Methods for Deep Reinforcement Learning".
MIT License
1.23k stars 279 forks source link

After some steps, all the NNs always output same action #75

Open Eify666666 opened 3 years ago

Eify666666 commented 3 years ago

I'm training a A3C these days, but the NN always take the same action, after some steps. The game I train for is similar to playing Go. There will be few reward in the short term. So it hard to learn something useful for the NN form the game. Maybe that is where the problem is. I tried ' torch.nn.utils.clip_grad_norm(lnet.parameters(), 50) ', and used relu as activate function. But it doesn't work.

RuoyuG commented 2 years ago

I meet same problem, it looks like stuck in a local optimal. Do you solve it?