search

araffin / learning-to-drive-in-5-minutes

Implementation of reinforcement learning approach to make a car learn to drive smoothly in minutes
https://towardsdatascience.com/learning-to-drive-smoothly-in-minutes-450a7cdb35f4
MIT License
284 stars 88 forks source link

Generate the same road #22

Closed ChunJyeBehBeh closed 4 years ago

ChunJyeBehBeh commented 4 years ago

Hello, I would like to train different RL algorithms (DDPG, PPO and SAC) using the same track in the simulator. I tried to go through the issues (#7 and https://github.com/tawnkramer/sdsandbox/issues/15) but still cannot generate the same track every time I start the simulator.

I even tried to hard code the seed in the code. But still cannot manage to solve it. May I know how to achieve it?

def seed(self, seed=None):
        self.np_random, seed = seeding.np_random(10)  #hardcode the seed to 10
        return [seed]

Update:

I found you have a pull request which have that functionality under the branch pretrain, it mean I need to checkout to that branch in order to achieve that?

araffin commented 4 years ago

mean I need to checkout to that branch in order to achieve that?

yes, you do have to use that branch, you also need a newest build of the simulator.

KarlXing commented 3 years ago

Hi @ChunJyeBehBeh @araffin , I also have the need to use the same track. Could you please give some insights for where should I look into? Thanks!

ChunJyeBehBeh commented 3 years ago

@KarlXing Do checkout the pretrain branch and there is one parameter 'seed' for you to set. Besides, you need to build latest simulator yourself. Let me know if you face any other issues.