AI, ML, RL

• Machine Learning

  • Supervised Learning: Task of learnning from labeled data

    ex) image recognition

  • Unsupervised Learning: Task of learning from unlabeled data

    ex) classification

  • Reinforced Learning: Task of learning through trial and error

    ex) game

RL Problem: A Formal Definition

complex sequential decision-making under uncertainty

• Complex: large state (information) and action (control option) fields
• Sequential: delayed consequences, and unknown relationship between action and next states
• Uncertainty: there are some randomness and noise

Agent and Environment

• Agent: decision maker
• Environment: the others (— the problem representation) → everything outside the agent

“Repeated” interactions

Goal: Teaching the agent how to behave by telling it how good it’s doing

1. observation
2. take an action
3. state changes according to the agent’s action
4. react with new observation and reward
5. repeat

Terminology

• Policy: agent’s mapping from state to action
  • $A_t=\pi(S_t)$
• Model (of environment): some known structure of env.
  • Some information about $p(S_{t+1}, R_{t+1}|S_t,A_t)$