Policy gradient
Also: policy gradient methods, REINFORCE, PPO, proximal policy optimization
The policy gradient theorem gives ∇J(θ) = E[∇ log π_θ(a | s) · G], so actions followed by high returns get their log-probability pushed up. REINFORCE uses sampled returns; baselines, actor-critic and PPO reduce its high variance.
∇_θ J(θ) = E_π[ ∇_θ log π_θ(a | s) · G_t ]
Example
If an episode where the agent chose "left" earns return G = +5 and another where it chose "right" earns −2, the update raises the probability of "left" and lowers "right" in that state.