* Add base for Soft Actor-Critic
* Pick changes from old SAC branch
* Update sac.py
* First implementation of sac model
* Remove unnecessary SAC imports
* Prune unnecessary noise and exploration code
* Implement SAC model and use that in SAC policy
* runs but doesn't learn
* clear state
* fix batch size
* Add missing alpha grads and vars
* -200 by 2k timesteps
* doc
* lazy squash
* one file
* ignore tfp
* revert done
* [rllib] Separate optimisers for DDPG actor & crit.
* [rllib] Better names for DDPG variables & options
Config changes:
- noise_scale -> exploration_ou_noise_scale
- exploration_theta -> exploration_ou_theta
- exploration_sigma -> exploration_ou_sigma
- act_noise -> exploration_gaussian_sigma
- noise_clip -> target_noise_clip
* [rllib] Make DDPG less class-y
Used functions to replace three classes with only an __init__ method & a
handful of unrelated attributes.
* [rllib] Refactor DDPG noise
* [rllib] Unify DDPG exploration annealing
Added option "exploration_should_anneal" to enable linear annealing of
exploration noise. By default this is off, for consistency with DDPG &
TD3 papers. Also renamed "exploration_final_eps" to
"exploration_final_scale" (that name seems to have been carried over
from DQN, and doesn't really make sense here). Finally, tried to rename
"eps" to "noise_scale" wherever possible.
* add marvil policy graph
* fix typo
* add offline optimizer and enable running marwil
* fix loss function
* add maintaining the moving average of advantage norm
* use sync replay optimizer for unifying
* remove offline optimizer and use sync replay optimizer
* format by yapf
* add imitation learning objective
* fix according to eric's review
* format by yapf
* revise
* add test data
* marwil
A bunch of minor rllib fixes:
pull in latest baselines atari wrapper changes (and use deepmind wrapper by default)
move reward clipping to policy evaluator
add a2c variant of a3c
reduce vision network fc layer size to 256 units
switch to 84x84 images
doc tweaks
print timesteps in tune status
Rename AsyncSamplesOptimizer -> AsyncReplayOptimizer
Add AsyncSamplesOptimizer that implements the IMPALA architecture
integrate V-trace with a3c policy graph
audit V-trace integration
benchmark compare vs A3C and with V-trace on/off
PongNoFrameskip-v4 on IMPALA scaling from 16 to 128 workers, solving Pong in <10 min. For reference, solving this env takes ~40 minutes for Ape-X and several hours for A3C.
