search

Kaixhin / Rainbow

Rainbow: Combining Improvements in Deep Reinforcement Learning
MIT License
1.56k stars 282 forks source link

Fix stddev default to match NoisyNet paper #40

Closed stormont closed 5 years ago

stormont commented 5 years ago

Small change to correct the default standard deviation initialization parameter for Factorized Noisy Nets.

For an out-of-the-box replication, this matches the published research.

References

(Fortunato et al. 2017) Noisy Networks for Exploration - Section 3.2, last sentence.

(Hessel et al. 2017) Rainbow: Combining Improvements in Deep Reinforcement Learning - Table 1

Kaixhin commented 5 years ago

Thanks (also thanks for including the sections within the papers, that made it really easy to check)!