Skip to content
New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Jump to bottom

Version 1.1.0 Classical / Modern RL Models #18

Open
wants to merge 31 commits into
base: master
Choose a base branch
from
Open

Version 1.1.0 Classical / Modern RL Models #18

wants to merge 31 commits into from

Conversation

josiahls
Copy link
Owner

@josiahls josiahls commented Mar 2, 2020 •
edited
Loading

Classical / Modern RL Models

  • Add Cross Entropy Method CEM
  • NStep Experience replay
  • Gaussian and Factored Gaussian Noise exploration replacement
  • Add Distributional DQN
  • Add RAINBOW DQN
  • Add REINFORCE
  • Add PPO
  • Add TRPO
  • Add D4PG
  • Add A2C
  • Add A3C
  • Add SAC

josiah and others added 30 commits February 2, 2020 12:00
Added:
fd08f83 
- initial updates to notebooks
Added:
8e5a7d0 
- fixed target lunar lander
Updated:
c06666c 
- notebooks
Fixed:
- there seems to be a strange versioning issue with whether the keyword axis needs to be passed on argmax pytorch functions
Added:
f28ae7f 
- fixed target lunar lander
Merge remote-tracking branch 'origin/version_1_0_0' into version_1_0_0
ac91d52
Note:
494efc1 
- there seems to be an overall issue with image generation
Added:
dc9671b 
- initial gifs, finished notebooks
Added:
76a633f 
- gif table generating notebook
Added:
9ca68f7 
- reward graphs
Added:
ec37505 
- reward graphs
Added:
2d3d466 
- reward graphs
Added:
006b41f 
- reward graphs
Added:
336796d 
- reward graphs
Added:
b1d7aa6 
- reward graphs
Added:
3a09d5f 
- initial TRPO step code. Highly likely this is way off. This is a first attempt as translating math of the research paper into a code implementation. Excited to see how close I was to the real implimentation
Added:
cb10788 
- first good start with REINFORCE
Added:
c44a6aa 
- REINFORCE is training now, but doesnt work. What happens when the actions are binary? An action with probability 1 is always going ot be sampled!
Added:
57d922c 
- cross entropy method. Does not seem to work great right now, pretty sure an existing bug in the code..
Removed:
- reinforce and trpo code. Need to start over...
Fixed:
5a3094b 
- OK THIS IS BIG, WEIGHT DECAY FUCKS THINGS. This might mean that other RL models might perform better also with weight decay set to 0...
Changed:
a1dfbad 
- current cem test is now flagged as a performance test.
Added:
cab3452 
- NStep Experience replay. Very very promising
Added:
0828544 
- guassian noise layers. They do not seem to improve performance on cartpole, but may do better on Atari games
- ROADMAP items
@josiahls
Merge branch 'master' into version_1_1_0
c3367ba
Fixed:
af78818 
- Greedy epsilon crashing lol
Fixed:
9a9595b 
- old reinforcement failing unit tests.
Changed:
bc1c2b9 
- resolution wrapper handles other returns from render better
Added:
- distributional dqn. Does not seem to work well with cartpole, investigating
Added:
deee785 
- alternate dist dqn, which trains quickly now
Added:
f30fe6b 
- RAINBOW dqn. Currently it is one of the worst performing dqns. In the next update (1.2)
Updated:
2c09b44 
- roadmap
Added:
2b62455 
- REINFORCE model for Cartpole
Updated:
fc6765c 
- REINFORCE roadmap
# for free to join this conversation on GitHub. Already have an account? # to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@josiahls