TD3+BC — the simplest offline RL baseline.

medium

Answer

$Fujimoto & Gu 2021.$
$({\pi}(s)\; - \;a_{\mathrm{data}})^{2}$
$Encourages policy to stay close to data actions.$
$Extremely simple (add one term), strong performance on D4RL.$
$Foundation demonstrating that a small BC regularization is often enough for offline RL — sophisticated methods (CQL, IQL) add marginal gains.$

Check yourself — multiple choice

Random
TD3 + BC regularizer λ $({\pi}(s)\; - \;a_{\mathrm{data}})^{2}$ added to policy loss; extremely simple + strong D4RL performance; showed a small BC term is often enough
Same as SAC
Not real

TD3+BC: add BC term to TD3 loss; simplest strong offline RL baseline.

#offline-rl

214 interview questions in this topic.