In a reinforcement learning setup, an autonomous agent interacts with an environment to learn an optimal policy. What is the term for the numerical feedback signal returned by the environment to evaluate the quality of the agent's action?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Think of reinforcement learning like training a puppy. If the puppy sits when you tell it to, you give it a treat. If it chews on your favorite shoes, you don't. In RL, the puppy is the "agent," and that treat is the "reward." It's a numerical value sent back by the environment after every action the agent takes. A positive number means "good job, do more of that," and a negative number means "stop doing that." The agent's entire goal in life is to maximize the total reward it gets over time. It's not a hyperparameter you set before training, and it's not the state of the environment itself—it's the feedback signal that drives the whole learning process.
Full explanation below image
Full Explanation
Reinforcement Learning (RL) is a machine learning paradigm where an agent learns to make decisions by interacting with an environment. The core components of RL are: - Agent: The decision-maker (model) that performs actions. - Environment: The world the agent interacts with. - State: The current situation or configuration of the environment. - Action: A move or choice made by the agent. - Reward: A scalar feedback signal (a numerical value) returned by the environment immediately after the agent takes an action.
The reward signal defines the goal in a reinforcement learning problem. If the action moves the agent closer to the goal, the reward is positive; if it moves it away or causes a failure, the reward is negative or zero. The objective of the agent is to learn a policy—a mapping from states to actions—that maximizes the cumulative reward (return) it receives over time.
- Option A is incorrect because a hyperparameter is a configuration setting set before training (like the learning rate or discount factor), whereas the reward is dynamic feedback received during training. - Option C is incorrect because the state represents the current environment conditions, not the feedback signal. - Option D is incorrect because the model output in RL is the action selected by the agent's policy, whereas the reward is the environmental evaluation of that action.