#gradient-ascent

Question 1: Single choice

Optimization can maximize an objective as well as minimize a loss. To maximize a differentiable objective $J(\theta)$ , such as an average reward, with gradient steps, how must the parameters move?

Along $+\nabla_\theta J$ , which is the same as minimizing $-J$ 1
Along $-\nabla_\theta J$ , exactly as for a loss2
Along $+\nabla_\theta J$ for weights and $-\nabla_\theta J$ for biases3
Nowhere; gradients only work for minimizing, so a gradient-free search is needed4

Timed test