#log-prob

Question 1: Single choice

mediumAI-checkedPolicy Gradient & Actor-Critic

A diagonal Gaussian policy for a 6-dimensional action space computes logp = Normal(mu, std).log_prob(act) for a batch of 64 actions, where mu, std and act all have shape (64, 6). What do you need for the policy-gradient loss?

logp has shape (64, 6); sum over the last dimension to get one log-probability per action, shape (64,)1
logp already has shape (64,), one value per action2
logp has shape (64, 6); average over the last dimension to get one value per action3
logp has shape (64, 6); take the maximum over the last dimension4

Timed test