DeepRL icon indicating copy to clipboard operation
DeepRL copied to clipboard

Torcs training

Open sanjeevk42 opened this issue 7 years ago • 9 comments

Can you share the training details(training parameters, algorithm, total training steps etc) for torcs? Also did you have issues related to entropy becoming zero while training with a3c?

sanjeevk42 avatar Aug 01 '17 07:08 sanjeevk42

You can see hyperparameters in the file below. https://github.com/only4hj/DeepRL/blob/master/env/torcs/torcs_env.py (See parameters with option args.drl == 'dqn') Main algorithm is DDPG. The train is done under 1,000 episodes (4 hours in my PC) when low features are used as the input directly instead of image pixels.

When I used a3c I couldn't see the zero entropy as you mentioned. Did you add entropy loss?

futurecrew avatar Aug 01 '17 13:08 futurecrew

Actually I was training torcs with a3c implementation (https://github.com/openai/universe-starter-agent) by openai. But it doesn't seem to work for torcs due to drop in entropy after few iterations. I thought you might have encountered similar issue with a3c.

Did you try training torcs with image data on a3c?

sanjeevk42 avatar Aug 02 '17 13:08 sanjeevk42

I have not tried using a3c for torcs. I used only DDPG for torcs.

futurecrew avatar Aug 02 '17 23:08 futurecrew

Did you try DDPG with image data only?

sanjeevk42 avatar Aug 02 '17 23:08 sanjeevk42

Yes I did. It works fine with DDPG using only image input. With low features input the network outputs three actions, steering, acceleration and brake. With only image input i made the network output one action, steering of the car to make traing simple.

futurecrew avatar Aug 02 '17 23:08 futurecrew

Do you have any plan to publish the pre-trained models for torcs with low dimensional or high dimensional features?

ellie-ba avatar Aug 04 '17 03:08 ellie-ba

For now there is no plan to publish trained models. You can install and train them by following the setup below. https://github.com/only4hj/DeepRL/blob/master/env/torcs/INSTALL.md

futurecrew avatar Aug 29 '17 22:08 futurecrew

Hi! I have been getting an error. I was hoping if you could help with this. It keeps printing the same output.

Timeout for client answer Timeout for client answer

uhvardhan avatar Mar 21 '18 07:03 uhvardhan

Hi! I have been getting an error. I was hoping if you could help with this. It keeps printing the same output.

Timeout for client answer Timeout for client answer Same problem and on Torcs's graphical interface, the vehicle does not move very much. After a certain period of time (about 10 minutes), the vehicle bumps into the roadside (I suspect this is because the road is inclined and the car slides to one side automatically).

I was wondering if you could offer me some possible ways to solve this problem. I would appreciate your reply.

hilbertletanger avatar Mar 12 '19 10:03 hilbertletanger