Parameterised Policy

reinforcement-learning

Definition

Parameterised Policy

A parameterised policy ${\pi }_{\theta }$ is a stochastic policy whose action probabilities are computed by a function with parameters $\theta \in {\mathbb{R}}^d$:

$${\pi }_{\theta }(a\mid s)=\mathrm{Pr}(a_t=a\mid s_t=s;\theta ).$$

Common parameterisations include a neural network that outputs logits over $\mathcal{A}$ followed by a softmax (discrete actions) or outputs the mean and variance of a Gaussian (continuous actions). The parameters $\theta$ are updated via gradient ascent on a policy gradient objective.