search

garlicdevs / Fruit-API

A Universal Deep Reinforcement Learning Framework
http://fruitlab.org/
GNU General Public License v3.0
67 stars 22 forks source link

what is reward_clip_thresholds?? #6

Closed TyrandeWhisperwind closed 4 years ago

TyrandeWhisperwind commented 4 years ago

i would like to know please, what is reward_clip_thresholds? and why it is set to None in MODQNLearner(DQNLearner)? also do u have the algorithm to share with us of the MODQNLearner, so we can understand it better, cause in ur code their is no comments which makes it harder to guess how it works, thanks in advance

TyrandeWhisperwind commented 4 years ago

also where can i specify the weight of objectifs in MODQN?

garlicdevs commented 4 years ago

Hi, You can see the example of using MODQN here https://github.com/garlicdevs/Fruit-API/blob/master/fruit/samples/basic/multi_objectives_test.py

reward_clip_thresholds is used to clip the rewards if out of bound

garlicdevs commented 4 years ago

I implemented those based on this paper https://arxiv.org/abs/1803.02965

TyrandeWhisperwind commented 4 years ago

is it possible not to clip the rewards?

garlicdevs commented 4 years ago

Yes, just specify None

TyrandeWhisperwind commented 4 years ago

and if i want to add a new environment to solve my issue what should be changed in order to do so?

garlicdevs commented 4 years ago

In the example code, I implemented Deep Sea Treasure (DST) environment, you just check that code and develop your environment to replace the DST, other code can be reused

TyrandeWhisperwind commented 4 years ago

thank you very much!!