Reinforcement learning (RL) studies how an agent can learn to act in an environment to maximize rewards. Unlike methods trained from labeled examples, an RL agent learns through interaction and feedback.
In supervised learning, a learner uses labeled examples; in unsupervised learning, it looks for structure in unlabeled data. In RL, an agent repeatedly acts in an environment and receives feedback in the form of rewards. The agent's goal is not simply to predict a label, but to learn a behavior that produces high cumulative reward.
All flashcards
What is reinforcement learning?
A machine-learning approach in which an agent learns how to act by interacting with an environment and using reward feedback to improve its behavior.
What are the four core components of an MDP?
The state space, action space, transition probabilities, and rewards.
What is a policy?
A rule that specifies which action an agent should take in a state. A stochastic policy gives probabilities to available actions.
What does partial observability mean?
The agent cannot directly access the complete current state, perhaps because information is missing or observations are noisy. Such problems are commonly represented as POMDPs.
Why does an RL agent need rewards?
Rewards provide feedback about transitions and help the agent learn behavior that maximizes cumulative outcomes.
Quiz questions
In an MDP, what does the reward R_a(s,s') represent?Answer: The immediate feedback associated with that state transition.
What does a stochastic policy specify for a state?Answer: The probability of selecting each available action.
How do basic Monte Carlo methods estimate values?Answer: By averaging returns from complete episodes.
How can temporal-difference methods update value estimates?Answer: They use a transition and an estimate of value at the next state.