Policy (RL)
Also: policy, π(a|s), policy network
A policy π maps states to actions: deterministic a = π(s), or stochastic π(a | s), a probability for each action. Solving an RL problem means finding an optimal policy π* that maximizes expected return from every state.
π(a | s) = P(A_t = a | S_t = s)
Example
A stochastic gridworld policy might choose "right" with probability 0.7 and "up" with 0.3 in a cell next to the goal; the greedy policy from a learned Q-table just picks the action with the largest Q(s, a).