3.2.1
What Is a Policy in Reinforcement Learning?
Understand a policy as a mapping from states to action probabilities, including deterministic and stochastic policies.
What is it?
A policy defines how an agent selects actions in a given state. It is represented by π(a|s), describing the probability of selecting action a in state s. Policies can be deterministic or stochastic.
Key points
1. Policy determines action-selection behavior.
2. π(a|s) represents action probability.
3. Deterministic policy: a = π(s).
4. Stochastic policy: a ~ π(·|s).
5. Policy guides agent interaction.
6. Reinforcement learning improves the policy to increase expected return.
Example: Choosing a route
An autonomous delivery agent reaches a road junction.
Deterministic policy: Always turn left.
Stochastic policy: Turn left with probability 0.7 and right with probability 0.3.
Both define how the agent selects an action.
Real-world application
Policies are used in autonomous navigation, games, robotic control and AI decision-making.
Key takeaway
A policy is the agent’s decision-making rule for selecting actions in different states.