Description

---

Reimplementation of Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor.

Contributions are welcome. If you find any mistake (very likely) or know how to make it more stable, don't hesitate to send a pull request.

Requirements

---

• mujoco-py
• Plotly
• PyTorch

Run

---

For SAC :

python main.py --env-name Humanoid-v2 --scale_R 20 

For SAC (Hard Update):

python main.py --env-name Humanoid-v2 --scale_R 20 --tau 1 --target_update_interval 1000

For SAC (Deterministic, Hard Update):

python main.py --env-name Humanoid-v2 --scale_R 20 --deterministic True --tau 1 --target_update_interval 1000

Results

---

My results on Humanoid-v2 environment using SAC, SAC(hard update) and SAC(deterministic, hard update). This is a plot of average rewards at every 10000 step interval

Parameters

---

Parameters	Value
Shared	-
optimizer	Adam
learning rate(--lr)	3x10−4
discount(--gamma) (γ)	0.99
replay buffer size(--replay_size)	1x106
number of hidden layers (all networks)	2
number of hidden units per layer(--hidden_size)	256
number of samples per minibatch(--batch_size)	256
nonlinearity	ReLU
SAC	-
target smoothing coefficient(--tau) (τ)	0.005
target update interval(--target_update_interval)	1
gradient steps(--updates_per_step)	1
SAC (Hard Update)	-
target smoothing coefficient(--tau) (τ)	1
target update interval(--target_update_interval)	1000
gradient steps (except humanoids)(--updates_per_step)	4
gradient steps (humanoids)(--updates_per_step)	1

Environment	Reward Scale
HalfCheetah-v2	5
Hopper-v2	5
Walker2d-v2	5
Ant-v2	5
Humanoid-v2	20