AlgoPlus//machine learning / policy-gradient
Read the theory

Policy Gradient (REINFORCE)

Optimize the policy directly by pushing up the probability of rewarding actions.

Iteration
Reward
Legend
Action probability
Sampled action
AI Tutor Workspace
In a nutshell
Policy gradient learns the behaviour directly instead of first estimating how good each state is. The policy is a set of probabilities over the actions; the agent samples an action, sees the reward, then adjusts the probabilities — gradient ascent — to make rewarding actions more likely and poor ones less likely. Over many samples the policy drifts toward whatever earns the most reward. The classic version of this is the REINFORCE algorithm.
Ready
Press play to begin the cinematic walkthrough.
Skip the bookkeeping of values — just do more of what works. Roll the dice on an action, and if it paid off, dial up its probability a little next time.
Key terms
Go deeper in the lesson
Read the full theory, intuition & complexity for Policy Gradient (REINFORCE).