Convergence problems with a3c

[diyor...@gmail.com]

Hello, I've built an A3C implementation in keras using this as referance: /https://jaromiru.com/2017/03/26/lets-make-an-a3c-implementation/

And I'm using custom environment, which I have tested before on DQN and it sucessfully converged showing really good results. But when I use the same environment in A3C, it results in model just choosing the same action over and over. I tried changing some hyper-parametrs, but no result. I also tried using target model and updating it every n episodes, which resulted in better convergence in case of gym CartPole environment, buy still no effect on performance of my model in my custom environment. Any ideas are welcomed, thank you.

[lesley...@gmail.com]

Hi, I'm having the same issue in a3c, did you solve it by anyway?

Thanks

在 2018年6月21日星期四 UTC+9下午5:24:50，diyor...@gmail.com写道：