← AI Terminology
SAC - Soft Actor-Critic
SAC (Soft Actor-Critic) is an off-policy actor-critic RL algorithm that maximises expected return plus policy entropy, encouraging stable, exploratory continuous control.
A default strong baseline for continuous robotics tasks.
A default strong baseline for continuous robotics tasks.
Why It Matters in AI
Continuous control needs stable off-policy learners. SAC’s maximum-entropy objective improved robustness and became a go-to algorithm in MuJoCo-style benchmarks and robot learning stacks.
Key Points
| Aspect | Description |
|---|---|
| Libs | Stable-Baselines3, RLlib |
| Pros | Sample efficient; stable relative to older methods |
| Type | Off-policy actor-critic; continuous actions |
| Origin | Haarnoja et al., Soft Actor-Critic |
| Entropy | Temperature balances reward vs randomness |
| Related | TD3, PPO (on-policy alternative) |
Simple Analogy
A explorer paid both for treasure and for keeping diverse routes open — success plus healthy randomness.
Common Usage Examples
sb3.SAC('MlpPolicy', env)- MuJoCo locomotion training
- Tune entropy temperature
- Compare to PPO on the same robot task
Summary
In short: SAC is a maximum-entropy off-policy actor-critic algorithm — a standard choice for continuous-control reinforcement learning.