policy

A policy is the agent's brain in one word: its rule for choosing what to do. Given the current state, the policy says which action to take — or, more generally, gives a probability to each possible action. Everything the agent ever does flows from its policy; learning, in reinforcement learning, ultimately means improving this rule.

It is written pi, with pi of a given s meaning the probability of action a in state s. A policy can be as simple as a lookup table for a small grid world, or as complex as a deep neural network mapping camera pixels to motor torques. What matters is the mapping from situations to behaviour. Two big distinctions follow: a policy can be deterministic or stochastic, and it can be the optimal one or merely a current best guess.