HI,I have a question about this project?

sufengniu / gym_torcs_rl

reinforcement learning playground

2 stars 2 forks source link

HI,I have a question about this project? #2

Open BCWang93 opened 5 years ago

BCWang93 commented 5 years ago

I see this project in this repositories.'gym_torcs_rl/testing/DDPG-Keras-Torcs/'.In this project ,I see you use image as the input for the network.I want to know.Do you use the pretrained 'guide network' to train actor or critic?On the other hand,you only use image as the input,but not conclude the 'guide network'.Because I also need to slove this has 'guide net' DDPG,but I have many difficult.So I want to need your help.Can you help me?Thanks! (Note: Well,I mean is this should have three network) Thanks!

sufengniu commented 5 years ago

@BCWang93 we originally found that the algorithm (DDPG) using image is not converged, Therefore, we first train a 'guide network' that use low dimensional feature (original repo), then use 'guide network' as supervision help to train the image based model. However, we found even using these auxiliary supervision, after switch back to RL training, the network will eventually failed. our project finally suspended since we haven't got any success to let the network to converge.

BCWang93 commented 5 years ago

@sufengniu Thanks for your reply!But I have another question.I want to know the project in your repositories use the image as the input is how to setting.I need the use image as the network input recently.Can you give some help for me?Thanks!In addition,your code use image as input has some errors!I had comment another issu. Can you help me to slove it?Thank you very much