Gym: Random Seed Feature

Created on 19 Jul 2016 · 4Comments · Source: openai/gym

I am just wondering if the user can set a random seed so it could reproduce all game states for testing purpose. Is it possible to support it if there is no such feature? Thanks!

Source

RuofanKong

Most helpful comment

Yes, envs in gym expose a seed() function for exactly this purpose e.g.

>>> import gym
>>> env = gym.make('CartPole-v0')
>>> env.seed(0)

This could be documented better.

jietang on 21 Jul 2016

👍4

All 4 comments

Yes, envs in gym expose a seed() function for exactly this purpose e.g.

>>> import gym
>>> env = gym.make('CartPole-v0')
>>> env.seed(0)

This could be documented better.

jietang on 21 Jul 2016

👍4

@jietang Thanks! I also wonder if env.reset() will reset the seed that I previously set. For example,

import gym
env = gym.make('CartPole-v0')
env.seed(0)

# reset the environment.
env.reset()
# Then the random seed is still 0 or not?

In this case, does reset() reset seed as well?

RuofanKong on 21 Jul 2016

Hopefully not, otherwise the agent would always train on the same scenario.

danijar on 31 Jul 2016

(Correct, it does not.)

On Sat, Jul 30, 2016 at 3:39 PM, Danijar Hafner [email protected]
wrote:

Hopefully not, otherwise the agent would always train on the same scenario.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/openai/gym/issues/250#issuecomment-236393754, or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAM7kSn_8X85UwEEKnVq_EkEQyfrAxeeks5qa9KmgaJpZM4JP_vq
.