Introduction
In supervised learning, every example comes with the right answer, so the loss tells the network exactly how to improve. In RL, the agent only gets rewards, often long after the actions that earned them. How can we compute gradients for a policy network?
Policy gradient methods answer with a simple idea: play some episodes, then make the actions that led to good outcomes more likely, and the actions that led to bad outcomes less likely. Weight each adjustment by how good the outcome was.
Think of a basketball player practising free throws. They don't know the exact correct muscle movements. They shoot, see whether the ball goes in, and gradually repeat more of the movements that preceded successful shots. That's a policy gradient, carried out by a human brain.