Improve parameter distributions used in TestGaussianDistribution

chainer / chainerrl

ChainerRL is a deep reinforcement learning library built on top of Chainer.

MIT License

1.18k stars 224 forks source link

Closed muupan closed 5 years ago

muupan commented 5 years ago

Before this PR, both mean and var is sampled uniformly from [0,1]. This is not good because

the case when mean has negative elements is not tested, and
variance close to zero can lead to numerical instability, leading to failure in tests. (https://ci.preferred.jp/chainerrl.py3.cpu/26601/)

This PR changes the parameter distributions so that mean is sampled from N(0,1) while variance is sampled from U(0.5,2).

I locally ran TestGaussian 100 times without failure.

muupan commented 5 years ago

/test

pfn-ci-bot commented 5 years ago

Successfully created a job for commit c91d7cc:

muupan commented 5 years ago

/test

pfn-ci-bot commented 5 years ago

Successfully created a job for commit d62fe4c: