= Reparameterization gradient
= Pathwise gradient
{synonym}
A <reparameterization gradient> differentiates an <expected value> by representing the simulated <random variable> as a differentiable transformation of parameter-independent noise. For $X_\eta=g_\eta(\varepsilon)$, the identity $\nabla_\eta\mathbb E[F_\eta(X_\eta)]=\mathbb E[\nabla_\eta F_\eta(g_\eta(\varepsilon))]$ requires justified <differentiation under the integral sign>. For a <normal distribution>, use $X=m+e^\rho\varepsilon$ with $\varepsilon$ drawn from the <standard normal distribution>.
Back to article page
