while I am in principle in favor of this change (I'd rather have prng state in the env than a separate global one) looks like there are more strings attached (judging by the test failures). Also putting @gdb in the loop - is there a reason to have a gym-global prng (vs local to environment)?

I think @joschu chose to do it this particular way; I'd originally had the prng always in the environment. I don't recall the exact reasoning.

Unfortunately, I don't remember the reasoning, so I'll leave it up to you @pzhokhov

okay then... @zuoxingdong, please fix the test failures, and we can merge.

@pzhokhov Previous failures are fixed. Could you help to check what is the current failure ?

gym/envs/algorithmic/tests/test_algorithmic.py .................         [  4%]
gym/envs/box2d/test_lunar_lander.py ..                                   [  5%]
gym/envs/tests/test_determinism.py FERROR: InvocationError for command '/usr/local/gym/.tox/py3/bin/pytest' (exited with code 1)
___________________________________ summary ____________________________________
ERROR:   py3: commands failed

I think the problem is lack of mujoco license key in the build... Which means tests on all external PR's are failing at the moment. Will fix shortly.

should be fixed now, please pull upstream master into your branch to update the PR

@pzhokhov Now it seems okay

@pzhokhov oops, the current error seems from the random seed

What I do now is to create a numpy random state as a class member within Space, do you think it can be a good idea to support an optional flag to seed the space object independently ?

Or provide a seed() API to the space, so that the user can seed the space object independently.

In specific either

    def __init__(self, shape=None, dtype=None, seed=0):
        import numpy as np # takes about 300-400ms to import, so we load lazily
        self.shape = None if shape is None else tuple(shape)
        self.dtype = None if dtype is None else np.dtype(dtype)
        
        self.np_random = np.random.RandomState(seed)

or

class Space(object):
...
    def seed(self, seed=0):
        self.np_random = np.random.RandomState(seed)

I am not sure I understand a use case for the seed() method in the API, could you explain?

@pzhokhov The motivation here is the user can set the environment and its space objects with different seeds.

@pzhokhov Now it seems works fine.

@pzhokhov Shall we merge this ?

@pzhokhov Conflicts resolved