Learning by rewards (RL made simple) quiz

Question 1

What best describes reinforcement learning (RL)?

Accepted Answer

An agent learning through rewards and penalties by interacting with an environment

Answer

Training a model using a large labeled dataset

Answer

Downloading correct answers from the internet

Answer

Memorizing a fixed set of rules given by the programmer

Question 2

In reinforcement learning, what is called an 'agent'?

Accepted Answer

The learner or decision-maker that observes and takes actions

Answer

The labeled dataset used for training

Answer

The reward score given at the end

Answer

The programmer who writes the RL code

Question 3

What does 'environment' mean in reinforcement learning?

Accepted Answer

Everything the agent interacts with, providing states and rewards

Answer

The reward signal the agent receives

Answer

The agent's internal memory of past actions

Answer

The programming language used to build the agent

Question 4

What role does a 'reward' play in reinforcement learning?

Accepted Answer

A numerical signal telling the agent how good or bad its action was

Answer

A prize given to the programmer when the agent improves

Answer

The number of training examples the agent has seen

Answer

A code that resets the game to the beginning

Question 5

What does the term 'state' mean in reinforcement learning?

Accepted Answer

A snapshot of the current situation the agent observes

Answer

The final score at the end of the game

Answer

The action the agent just chose

Answer

The list of all possible rewards in the environment

Question 6

In reinforcement learning, what is a 'policy'?

Accepted Answer

The agent's strategy for choosing which action to take in each state

Answer

The environment's fixed rulebook that the agent cannot change

Answer

The total number of training episodes completed

Answer

The penalty given to the agent for a wrong decision

Question 7

What is an 'episode' in reinforcement learning?

Accepted Answer

One complete run from start to a terminal state

Answer

A single reward signal received after one action

Answer

The total amount of time spent training

Answer

One individual action taken by the agent

Question 8

What is the 'exploration vs. exploitation' dilemma in reinforcement learning?

Accepted Answer

Balancing trying new actions versus sticking with known good actions

Answer

Choosing between using a CPU or GPU for training

Answer

Deciding whether to use supervised or unsupervised learning

Answer

Deciding how many layers to add to a neural network

Question 9

What does Q-learning do in reinforcement learning?

Accepted Answer

A method where the agent learns the value of each action in each state

Answer

A way to quickly label training data

Answer

An algorithm for finding shortest paths in databases

Answer

A tool for compressing images using artificial intelligence

Question 10

Why do we say reinforcement learning is based on 'trial and error'?

Accepted Answer

The agent tries actions, observes results, and keeps what works best

Answer

The agent reads a manual of correct actions before starting

Answer

The programmer provides the correct answer for every action

Answer

The agent searches the internet for solutions during training

Question 11

Why are 'delayed rewards' challenging in reinforcement learning?

Accepted Answer

Benefits from an action that only appear much later in the episode

Answer

Rewards that arrive too fast for the agent to process correctly

Answer

Rewards given only at the very start of an episode

Answer

Rewards stored in a special compressed file format

Question 12

How is reinforcement learning used to teach robots to walk?

Accepted Answer

The robot receives rewards for progress and learns through many trials

Answer

Programmers write exact instructions for every possible leg movement

Answer

The robot copies a human's walking motion from a video perfectly

Answer

The robot uses GPS to follow a pre-planned walking path

Question 13

How did AlphaGo learn to play the board game Go better than any human?

Accepted Answer

It combined learning from human expert games with self-play reinforcement learning

Answer

Programmers manually entered the moves from every Go game ever recorded

Answer

AlphaGo searched the internet for winning Go strategies in real time

Answer

It followed a fixed set of rules written by Go grandmasters

Question 14

What is the difference between a positive reward and a negative reward (penalty) in RL?

Accepted Answer

Positive rewards encourage good actions; negative rewards discourage bad actions

Answer

Positive rewards come from humans; negative rewards are given automatically

Answer

They are the same thing stored in different places in memory

Answer

Positive rewards speed up training; negative rewards slow it down

Question 15

What is 'reward hacking' and why is it a problem in reinforcement learning?

Accepted Answer

When the agent earns high rewards through unintended loopholes instead of the real goal

Answer

When a hacker breaks into the system and steals the agent's reward score

Answer

When rewards are calculated incorrectly due to a software bug

Answer

When the agent refuses to take any action to avoid negative rewards

Question 16

What does a 'discount factor' do in reinforcement learning?

Accepted Answer

It controls how much the agent values future rewards compared to immediate ones

Answer

It reduces the agent's learning rate automatically over time

Answer

It removes negative rewards from the training data

Answer

It discounts the total number of training episodes required

Question 17

What is a 'value function' in reinforcement learning?

Accepted Answer

A function that estimates expected future reward from a given state

Answer

A function that counts the total number of actions taken in an episode

Answer

A function that checks whether the agent is still alive

Answer

A function that saves the agent's learned policy to a file

Question 18

What is 'deep reinforcement learning' and why is it powerful?

Accepted Answer

RL combined with deep neural networks to handle complex inputs like images

Answer

RL that only works with very small datasets

Answer

RL run on underwater computers

Answer

A type of RL that only works for games with deep storylines

Question 19

What is the epsilon-greedy strategy in reinforcement learning?

Accepted Answer

The agent picks a random action with probability ε and the best known action otherwise

Answer

The agent always picks the action with the highest reward seen so far

Answer

The agent alternates between the first and last action in its list

Answer

The agent skips any action whose reward is below epsilon

Question 20

What is multi-agent reinforcement learning (MARL) and what makes it more complex than single-agent RL?

Accepted Answer

RL where several agents interact, cooperate, or compete in the same environment

Answer

RL where a single agent receives multiple different reward signals at once

Answer

RL that runs on multiple computers one after another

Answer

RL where the policy is updated multiple times within a single step

Question 21

How does 'reinforcement learning from human feedback' (RLHF) help train AI chatbots like ChatGPT?

Accepted Answer

Human raters provide preference feedback that becomes a reward signal for the AI

Answer

The AI learns by training only on human-written code

Answer

Humans manually write all of the AI's responses in advance

Answer

The AI copies replies directly from human chat logs

Question 22

What is the 'sparse rewards' problem in reinforcement learning?

Accepted Answer

When rewards are given very rarely, making it hard for the agent to learn

Answer

When the agent receives too many rewards too quickly

Answer

When negative rewards outweigh positive ones throughout training

Answer

When the agent cannot store all the rewards it receives in memory

Question 23

Why do some people have ethical concerns about using reinforcement learning to control autonomous weapons?

Accepted Answer

Because reward signals cannot fully capture human values, risking harmful unintended behaviors

Answer

Because RL agents are too slow to be useful in combat situations

Answer

Because RL requires a reliable internet connection unavailable in combat zones

Answer

Because RL agents can only operate in video games, not the real world

Question 24

What is the 'sim-to-real' transfer problem in robotics reinforcement learning?

Accepted Answer

The challenge of making a policy trained in simulation work in the real physical world

Answer

The difficulty of converting RL code from one programming language to another

Answer

When a simulation runs in real time instead of faster than real time

Answer

Converting a real robot's sensor readings into a simulation format

Question 25

What is a Markov Decision Process (MDP) and why is it the foundation of RL?

Accepted Answer

A framework of states, actions, transitions, and rewards where the next state depends only on the current state

Answer

A type of deep learning model with many hidden layers

Answer

A database storing every possible reward the agent can earn

Answer

A method for compressing large training datasets efficiently

Question 26

What is curriculum learning in reinforcement learning?

Accepted Answer

Starting with easy tasks and gradually increasing difficulty during training

Answer

Teaching the agent to learn many completely different subjects all at once

Answer

Shuffling training episodes randomly to prevent the agent from overfitting

Answer

Using a school's official curriculum to design the reward function

Question 27

What is 'reward shaping' and why do RL engineers use it?

Accepted Answer

Adding intermediate rewards to guide the agent and speed up learning

Answer

Changing the shape of the neural network to better fit the reward function

Answer

Removing all negative rewards to make training more stable

Answer

Normalizing all rewards to values between 0 and 1

Question 28

Why is 'safe exploration' an important challenge in real-world reinforcement learning?

Accepted Answer

Preventing dangerous actions during real-world learning to avoid injury or damage

Answer

Making sure the agent's code does not crash the training server

Answer

Encrypting the reward function so it cannot be changed by outsiders

Answer

Limiting the agent to only receiving positive rewards during training

Question 29

Why are video games a popular environment for testing reinforcement learning algorithms?

Accepted Answer

Because games offer clear rewards, fast simulation, and no real-world consequences for failures

Answer

Because game companies pay AI researchers to use their games for testing

Answer

Because video games are the only environment where RL has ever worked

Answer

Because RL can only learn tasks that have a graphical screen interface

Question 30

What is 'self-play' in reinforcement learning and what makes it effective?

Accepted Answer

When the agent competes against copies of itself to continuously improve

Answer

When the agent records its own gameplay and watches it back like a replay

Answer

When the agent's programmers take turns manually controlling it

Answer

When the same agent is used for both training and evaluation without any changes

Test yourself