A robot in a simulator observes its position and surroundings, selects a movement action, then receives information about the next state and a reward.
Learn a strategy that improves expected cumulative reward over the route rather than only one immediate reward. Do not assume a complete list of correct labeled movements.
Choose TWO appropriate descriptions of this reinforcement learning setup.
Distinguish states, actions and environment rewards, and describe a policy as the action-selection strategy. The goal is expected cumulative reward rather than only immediate reward or memorized supervised labels.
Detailed explanation
An action can change future states and rewards. Greedy immediate reward and expected cumulative reward need not coincide.
The policy describes action selection. Consider immediate and future consequences together rather than only the next reward.
Actions affect subsequent states and rewards. This interaction differs from only reading input-output examples with correct labels.
A policy and an answer-label list are different. The reward signal defining this learning objective is not replaced by supervised labels.
Learning balances exploration of unknown outcomes with exploitation of known useful actions. Reusing past actions alone is not always sufficient.
Check official documentation
Compare the following points with the official documentation.
Map the scenario to the AWS definitions of state, action, environment, reward and policy. Conceptually compare an immediately rewarding route with a route chosen for cumulative reward, and distinguish exploration from exploitation.
Official references
What to confirm
Explain a policy as the action-selection strategy without confusing rewards, correct labels and policies. Better immediate reward alone does not establish achievement of the long-term objective.
Key points
- State and action
- Reward
- Policy
- Exploration and exploitation
Notes
- Environment: Official AIF-C01 Task 1.1 and AI/ML documentation review; no model training or production changes are performed.
- Compare the definitions with current AIF-C01 Task 1.1 and the linked primary sources.
- A conceptual example does not establish measured model accuracy or authorize external operations.
Foundation review
Separate the interaction roles
The agent takes actions and the environment provides resulting state and reward information. Training can use a simulator; a robot example does not imply operating physical hardware.
Separate immediate reward from the long-term goal
A short-term sacrifice can support a better long-term strategy. If the reward design misses the actual objective, a higher reward need not mean the desired behavior.
Distinguish exploration and exploitation
Exploration investigates unfamiliar outcomes; exploitation selects known useful actions. The word agent in reinforcement learning does not imply that all LLM tool use or agentic AI uses the same learning method.