Remember RL
Full name: Reinforcement Learning
Memory aid: Rewards guide Learning from actions.
Reward design can encourage unintended behavior.
Study area: AI & machine learning · Batch 5
Relevant study for: Azure AI Fundamentals AI-901 · Azure AI Apps and Agents Developer AI-103 · Azure AI Fundamentals (retired) AI-900 · Azure AI Engineer (retired) AI-102
Practice with exam flashcards →AI Models & Assistants / Knowledge library
Reinforcement Learning (RL)
Learning actions that maximize cumulative reward
Why learn Reinforcement Learning (RL)?
Useful for sequential decisions where an action changes later opportunities and outcomes. A reward must match the intended task closely enough that optimizing it produces useful behavior.
The problem it solves
Learns how to act when success depends on a sequence of decisions and rewards may arrive after the actions that caused them.
Where it came from
Reinforcement learning trains a policy to select actions that maximize expected cumulative reward. Experience contains observations, actions, rewards, and subsequent observations, gathered through interaction or previously recorded trajectories. Delayed consequences make credit assignment and the choice of reward central to the problem.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Specify observations, available actions, and rewards for a small grid-world task.
- Calculate the discounted return of a three-step reward sequence.
- Implement an epsilon-greedy action rule for a small tabular policy.
- Compare a learned policy with a random policy across several evaluation seeds.
- Construct a reward loophole and revise the task specification to address it.
Documentation and sources
- Spinning Up: key concepts in reinforcement learning (spinningup.openai.com)
- Gymnasium: the agent-environment loop (gymnasium.farama.org)