[WIP] ReinforcementLearning.jl integration #9
Conversation
Codecov Report
@@ Coverage Diff @@
## main #9 +/- ##
==========================================
- Coverage 92.41% 92.31% -0.10%
==========================================
Files 81 81
Lines 3823 3761 -62
==========================================
- Hits 3533 3472 -61
+ Misses 290 289 -1
Continue to review full report at Codecov.
|
rejuvyesh
force-pushed
the
jkg/rlbase
branch
2 times, most recently
from
22e4549
to
b606aa1
Compare
examples/deeprl/cartpole_ppo.jl
Outdated
| actor = Chain( | ||
| Dense(ns, 256, relu; init = glorot_uniform(rng)), | ||
| Dense(256, na; init = glorot_uniform(rng)), | ||
| ), |
There was a problem hiding this comment.
Note that you are using the discrete version of PPO here. But the cart pole env here seems to be a continuous version. (The actions space is [-1.0, 1.0]). So you may take reference from https://github.com/JuliaReinforcementLearning/ReinforcementLearning.jl/blob/935f68b6cb378f9929a8d9914eb388e86213c86d/src/ReinforcementLearningExperiments/deps/experiments/experiments/Policy%20Gradient/JuliaRL_PPO_Pendulum.jl#L43-L50
There was a problem hiding this comment.
Good point! Thanks for checking in. Although currently I also need to define the reward/cost function for cartpole on Dojo side.
|
We should probably rethink the interface to ReinforcementLearning.jl once their updates are done (JuliaReinforcementLearning/ReinforcementLearning.jl#614) |
I realized that
CommonRLInterface.jlnever settled on what to do with continuous action spaces, so directly integrating with RLBase from ReinforcementLearning.jl.Will add tests and examples with PPO and DDPG.