Untitled Flashcards Set

1. Reinforcement Learning (RL)

Definition: RL is teaching a machine to learn by trial and error, using rewards (good) and penalties (bad) to improve decisions.
Example: A robot learns to exit a maze by getting rewards for correct moves.

2. Q-Learning

Definition: Q-learning is an RL method where an agent learns a Q-table to decide the best actions in each situation.
Example: A game character uses the Q-table to decide whether to jump or crouch.

3. Doing RL by Hand

Definition: Manually update Q-values by:

1. Picking actions (random/best).

2. Observing rewards.

3. Repeating to improve decisions.
   Example: Calculating the best moves in a small board game.

4. Bandits

Definition: A scenario where the agent chooses between options (like slot machines) to find the best reward.
Example: Picking the best slot machine in a casino.

5. Multi-Armed Bandits

Definition: Choosing between multiple options (e.g., slot machines), balancing trying new options (exploration) and using the best one (exploitation).
Example: Deciding which slot machine pays the most.

6. Contextual Bandits

Definition: Bandits with context—the choice depends on the situation.
Example: Showing ads based on user preferences.

7. Epsilon-Greedy Strategy

Definition: A method to balance exploration and exploitation:

• Explore randomly with chance ε.

• Exploit the best-known action otherwise.
  Example: Testing a new slot machine 10% of the time, using the best 90%.

8. Upper Confidence Bound (UCB)

Definition: A strategy balancing good rewards and uncertain actions.
Example: Testing underused slot machines while sticking to the best one.

9. Thompson Sampling

Definition: Uses probabilities to pick actions with the highest potential reward, balancing exploration and exploitation naturally.
Example: Sampling different slot machines to choose the most promising one.