Skip to content

Implementation for paper "A Consciousness-Inspired Planning Agent for Model-Based Reinforcement Learning".

License

Notifications You must be signed in to change notification settings

mila-iqia/Conscious-Planning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

By Mingde "Harry" Zhao, Zhen Liu, Sitao Luan, Shuyuan Zhang, Doina Precup and Yoshua Bengio

Install Dependencies

pip install -r requirements.txt

Reproducing Results

CP

python run_distshift_randomized_mp.py --method DQN_CP --num_explorers 8 --ignore_model 0 --disable_bottleneck 0 --size_bottleneck 8

UP

python run_distshift_randomized_mp.py --method DQN_CP --num_explorers 8 --ignore_model 0 --disable_bottleneck 1

WM

python run_distshift_randomized_mp.py --method DQN_WM --num_explorers 8 --ignore_model 0 --disable_bottleneck 0 --size_bottleneck 8 --period_warmup 1000000

Dyna

python run_distshift_randomized_mp.py --prioritized_replay 0 --method DQN_Dyna --num_explorers 8 --ignore_model 0 --disable_bottleneck 0 --size_bottleneck 8 --learn_dyna_model 1

Special thanks to my colleague and friend Safa Alver @alversafa for pointing out that Dyna should not use prioritized buffer as it shouldn't prioritize on the errors generated by potentially inaccurate imagined transitions, as well as the runtime bugs surrounding this matter!

Dyna*

python run_distshift_randomized_mp.py --method DQN_Dyna --num_explorers 8 --ignore_model 0 --disable_bottleneck 0 --size_bottleneck 8 --learn_dyna_model 0

NOSET

python run_distshift_randomized_mp.py --method DQN_NOSET --num_explorers 8 --ignore_model 0 --layers_model 2 --len_hidden 256

Changing Settings

Read run_distshift_randomized_mp.py!

About

Implementation for paper "A Consciousness-Inspired Planning Agent for Model-Based Reinforcement Learning".

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages