search

yanpanlau / DDPG-Keras-Torcs

Using Keras and Deep Deterministic Policy Gradient to play TORCS
717 stars 267 forks source link

Track change failed #6

Open Artioss opened 7 years ago

Artioss commented 7 years ago

I trained a model in the track CG Speedway number 1, and it work well when test in this track. But when I change the track(e.g. to CG track 2), and the trained model seem not work, the car run out of the track. Is that means for every different track, I have to train different models ?

yanpanlau commented 7 years ago

No. It probably means you are overfit the model in CG Speedway. I suggest you try following

1) Train on a more difficult track and test out in CG speedway number 1 instead. 2) Apply L2 Regularization to your Neural Network and make sure not overfit 3) Do cross-validation on training/testing different tracks.

In fact, even in the original Deepmind paper, they did not test their model on the validation data. I did validate my model where the training data and validation dataset are different.

Artioss commented 7 years ago

@yanpanlau Thanks for your suggestions, i'll try it soon. Could you please tell me what's your training & validation tracks? By the way, what's your opinion with issue #2? Thank you.

deepmeng commented 7 years ago

would you like tell us how you change the track by which interface or script? @yanpanlau @Artioss

yanpanlau commented 7 years ago

Hi,

1) sudo torcs 2) Click Race 3) Click Practice 4) Click Configure Race 5) Choose different track 6) Click New Race

Then you should see the game load a new track. Then you can click Ctrl + C to quit and run your AI agent.

On Wed, Apr 19, 2017 at 11:24 PM, Benjamin notifications@github.com wrote:

would you like tell us how you change the track by which interface or script? @yanpanlau https://github.com/yanpanlau @Artioss https://github.com/Artioss

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/yanpanlau/DDPG-Keras-Torcs/issues/6#issuecomment-295309095, or mute the thread https://github.com/notifications/unsubscribe-auth/AO1sY8x5C9XrB9Zx3oqc53FaQiS-kO87ks5rxicfgaJpZM4K-u6u .

deepmeng commented 7 years ago

Hi man, sorry to bother you, this visualization method we all know. Do you know is there a method to config it in some functions? I read and modify the *.py files, but it does not work.

yanpanlau commented 7 years ago

I have not do it myself before. The original author (gym_torcs) use a unix tools called xautomation to do the "actual mouse click", you may want to try that

On Wed, Apr 19, 2017 at 11:44 PM, Benjamin notifications@github.com wrote:

Hi man, sorry to bother you, this visualization method we all know. Do you know is there a method to config it in some functions? I read and modify the *.py files, but it does not work.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/yanpanlau/DDPG-Keras-Torcs/issues/6#issuecomment-295316658, or mute the thread https://github.com/notifications/unsubscribe-auth/AO1sY2zFfbjMLfhB8O47IopM-es41d8Eks5rxivwgaJpZM4K-u6u .

deepmeng commented 7 years ago

thanks very much!

deepmeng commented 7 years ago

Hi man, excuse me, could you please tell us how we can change the mode from "practice" to "quick race", and always keep it in "quick race"? THX very much!

JiaojiaoYe1994 commented 6 years ago

Hey, I currently also look into this project, have you found how to change the mode and could you please tell me? @deepmeng

deepmeng commented 6 years ago

@JiaojiaoYe1994 I'm very sorry about this. I can't remember it anymore...