How to train DDPG episode reward more better?

4 次查看（过去 30 天）

hunson yang 2020-2-26

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/507677-how-to-train-ddpg-episode-reward-more-better

评论： Guoge Tan 2020-5-25

I'm training a DDPG agent from the Reinforcement Learning toolbox. But as you can see, my episode reward never change. I try so many way to fix this problem. Like change the netwoek, Gradient Threshold, Learning Rate. But the result will be the same. I check my reward funtion, if the situation is eligible I will give it some reward or penalty. But its reward is always be same.

Is my condtion have some problem? Or my results are not input into the model? I dont have anyway to do.

2 个评论
显示无隐藏无

Emmanouil Tzorakoleftherakis 2020-2-28

How did you set the IsDone flag? This may lead to premature episode termination

Guoge Tan 2020-5-25

Hi, sorry to bother you, but I'd like to ask if your problem is solved or not? I‘m working on a path planning problem using the Reinforcement Learning toolbox on MATLAB R2020a and I also encountered a problem similar to yours.