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

The idea in plain language.

Learn useful behavior while preserving enough action diversity to keep exploring.

How it works

The actor learns a probability distribution over actions. Its objective balances expected reward with entropy, encouraging a policy that remains diverse while succeeding at the task. A critic estimates action values, and off-policy updates reuse previously collected experience. This combination is intended to improve data efficiency and reduce the instability of deep reinforcement learning.

What to keep in mind

The paper's continuous-control results are benchmark evidence, not a guarantee of safe behavior or robust learning in an arbitrary physical environment.

Source: Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.