search

p-christ / Deep-Reinforcement-Learning-Algorithms-with-PyTorch

PyTorch implementations of deep reinforcement learning algorithms and environments
MIT License
5.53k stars 1.19k forks source link

DDPG Acting Deterministically #9

Open JohnBurden opened 5 years ago

JohnBurden commented 5 years ago

Hi,

I've been using the DDPG implementation on the pyBullet implementation of RaceCar (https://github.com/bulletphysics/bullet3/blob/master/examples/pybullet/gym/pybullet_envs/bullet/racecar.py) after wrapping for compatibility with this repo. When running some basic experiments (essentially the same as those in the Results directory) I noticed that DDPG appears to be acting deterministically. Every run is the same when I set config to handle multiple runs. I do not believe this should be the case as the environment itself has an element of stochasticity in how it generates goals, and also I think the DDPG learning process is stochastic also. Is this possibly indicative of some random number seeding issue within this repo or how multiple runs are being handled?

Thanks, John

p-christ commented 5 years ago

Hi,

Ah yes, at the moment if you choose the multiple runs option it will do multiple runs with the same random seed that you set. So that might be explaining what's happening - do you see different results if you change the random seed?

Thanks,

Petros