Soft Actor-Critic

[kengz]

Feature / Fix

• implement Soft Actor-Critic (SAC) from https://arxiv.org/abs/1801.01290
• split and update roboschool benchmark specs
• update random baselines for roboschool
• change continuous policy std sampling from softplus to log-clamp-exp transform. verified working on PPO BipedalWalker
• add discrete control to SAC using reparametrizable Categorical distribution (Gumbel-Softmax)

Roboschool (continuous control) Benchmark

Note that the Roboschool reward scales are different from MuJoCo's.

Env. \ Alg.	A2C (GAE)	A2C (n-step)	PPO	SAC
RoboschoolAnt				1153.87 graph
RoboschoolHalfCheetah				1204.68 graph
RoboschoolHopper				1161.24 graph
RoboschoolWalker2d				695.36 graph

LunarLander (discrete control) Benchmark

Trial graph	Moving average

[CarloLucibello]

Hi, just wanted to point out, the follow-up paper by the authors of SAC
https://arxiv.org/abs/1812.05905
One of the main differences with the original paper is that they don't use a separate V network

[kengz]

Hi, just wanted to point out, the follow-up paper by the authors of SAC
https://arxiv.org/abs/1812.05905
One of the main differences with the original paper is that they don't use a separate V network

@CarloLucibello Thanks for pointing that out. This PR implements the first version of SAC and adds a discrete control version. I'll implement the improved version in the next PR.