1% of exploratory actions vs 10% for DQN

google / dopamine

Dopamine is a research framework for fast prototyping of reinforcement learning algorithms.

https://github.com/google/dopamine

Apache License 2.0

10.42k stars 1.36k forks source link

1% of exploratory actions vs 10% for DQN #131

Open Wovchena opened 4 years ago

Wovchena commented 4 years ago

Hi baselines/README.md states:

Unsurprisingly, DQN performs much better when trained with 1% of exploratory actions instead of 10% (as used in the original Nature paper).

Why is that unsurprising? Is there an explanation why DQN should be trained with 1% of exploratory actions instead of 10%? Or is this is just an empirical result which appears for most of DQN implementations?