Q#1: Reinforcement learning (RL) is:
(A) Learning to make sequential decisions through trial and error
(B) Learning from labeled examples only
(C) Learning from rules only
(D) BFS only
Answer: (A) Learning to make sequential decisions through trial and error
Q#2: In RL, the agent interacts with:
(A) The environment
(B) Training data only
(C) Decision trees only
(D) BFS only
Answer: (A) The environment
Q#3: A policy in RL defines:
(A) Mapping from states to actions
(B) Rewards only
(C) State transitions only
(D) BFS only
Answer: (A) Mapping from states to actions
Q#4: The reward function specifies:
(A) Immediate feedback for actions taken
(B) State transitions only
(C) Policies only
(D) BFS only
Answer: (A) Immediate feedback for actions taken
Q#5: The goal of an RL agent is to:
(A) Maximize cumulative reward over time
(B) Minimize immediate reward
(C) BFS only
(D) DFS only
Answer: (A) Maximize cumulative reward over time
Q#6: Markov Decision Process (MDP) is defined by:
(A) States, actions, transition model, and reward function
(B) States only
(C) Rewards only
(D) BFS only
Answer: (A) States, actions, transition model, and reward function
Q#7: Transition model in MDPs represents:
(A) Probability of next state given current state and action
(B) Reward only
(C) BFS only
(D) DFS only
Answer: (A) Probability of next state given current state and action
Q#8: Discount factor in RL controls:
(A) Importance of future rewards relative to immediate rewards
(B) Learning rate only
(C) BFS only
(D) DFS only
Answer: (A) Importance of future rewards relative to immediate rewards
Q#9: Value function in RL represents:
(A) Expected cumulative reward from a state under a policy
(B) Immediate reward only
(C) BFS only
(D) DFS only
Answer: (A) Expected cumulative reward from a state under a policy
Q#10: Action-value function (Q-function) represents:
(A) Expected return for taking an action in a state under a policy
(B) State only
(C) Reward only
(D) BFS only
Answer: (A) Expected return for taking an action in a state under a policy
Q#11: Bellman equation is used to:
(A) Relate value of a state to values of successor states
(B) Randomly assign values
(C) BFS only
(D) DFS only
Answer: (A) Relate value of a state to values of successor states
Q#12: Value iteration algorithm:
(A) Iteratively updates value function until convergence
(B) Random guesses
(C) BFS only
(D) DFS only
Answer: (A) Iteratively updates value function until convergence
Q#13: Policy iteration alternates between:
(A) Policy evaluation and policy improvement
(B) Random actions
(C) BFS only
(D) DFS only
Answer: (A) Policy evaluation and policy improvement
Q#14: Model-based RL uses:
(A) Known transition and reward functions to plan
(B) Only trial-and-error learning
(C) BFS only
(D) DFS only
Answer: (A) Known transition and reward functions to plan
Q#15: Model-free RL learns:
(A) Value functions or policies directly from experience
(B) From complete transition models
(C) BFS only
(D) DFS only
Answer: (A) Value functions or policies directly from experience
Q#16: Monte Carlo methods in RL:
(A) Estimate value functions based on complete episodes
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Estimate value functions based on complete episodes
Q#17: Temporal Difference (TD) learning:
(A) Updates estimates based on partial sequences
(B) Requires complete episodes
(C) BFS only
(D) DFS only
Answer: (A) Updates estimates based on partial sequences
Q#18: Q-learning is:
(A) A model-free off-policy RL algorithm
(B) BFS only
(C) DFS only
(D) Model-based only
Answer: (A) A model-free off-policy RL algorithm
Q#19: SARSA is:
(A) A model-free on-policy RL algorithm
(B) BFS only
(C) DFS only
(D) Off-policy only
Answer: (A) A model-free on-policy RL algorithm
Q#20: Exploration in RL ensures:
(A) Trying different actions to discover rewards
(B) Exploiting known actions only
(C) BFS only
(D) DFS only
Answer: (A) Trying different actions to discover rewards
Q#21: Exploitation in RL refers to:
(A) Selecting best-known actions for reward
(B) Random actions only
(C) BFS only
(D) DFS only
Answer: (A) Selecting best-known actions for reward
Q#22: ε-greedy strategy:
(A) Chooses random action with probability ε, best action otherwise
(B) Always selects best action
(C) BFS only
(D) DFS only
Answer: (A) Chooses random action with probability ε, best action otherwise
Q#23: Decaying ε in ε-greedy helps to:
(A) Shift from exploration to exploitation over time
(B) Always explore randomly
(C) BFS only
(D) DFS only
Answer: (A) Shift from exploration to exploitation over time
Q#24: Reward shaping is used to:
(A) Guide learning by providing intermediate rewards
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Guide learning by providing intermediate rewards
Q#25: Function approximation in RL is used when:
(A) State space is too large to represent values exactly
(B) State space is small
(C) BFS only
(D) DFS only
Answer: (A) State space is too large to represent values exactly
Q#26: Deep Q-Networks (DQN) combine:
(A) Q-learning with neural networks
(B) BFS only
(C) DFS only
(D) Model-based planning only
Answer: (A) Q-learning with neural networks
Q#27: Eligibility traces in RL combine:
(A) TD learning and Monte Carlo methods
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) TD learning and Monte Carlo methods
Q#28: On-policy methods learn:
(A) Value of policy being executed
(B) Value of another policy
(C) BFS only
(D) DFS only
Answer: (A) Value of policy being executed
Q#29: Off-policy methods learn:
(A) Value of optimal policy independent of actions taken
(B) Only current policy
(C) BFS only
(D) DFS only
Answer: (A) Value of optimal policy independent of actions taken
Q#30: Exploration-exploitation trade-off balances:
(A) Learning new knowledge vs. maximizing rewards
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Learning new knowledge vs. maximizing rewards
Q#31: Monte Carlo control learns:
(A) Optimal policy using returns from sampled episodes
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Optimal policy using returns from sampled episodes
Q#32: Policy gradient methods learn:
(A) Parameterized policies directly
(B) Value functions only
(C) BFS only
(D) DFS only
Answer: (A) Parameterized policies directly
Q#33: Actor-critic methods combine:
(A) Policy representation (actor) and value function (critic)
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Policy representation (actor) and value function (critic)
Q#34: Reward signal in RL is:
(A) Sparse, delayed, or noisy
(B) Deterministic always
(C) BFS only
(D) DFS only
Answer: (A) Sparse, delayed, or noisy
Q#35: Model-based planning in RL uses:
(A) Learned or given models of environment to plan actions
(B) Trial-and-error only
(C) BFS only
(D) DFS only
Answer: (A) Learned or given models of environment to plan actions
Q#36: Monte Carlo tree search (MCTS) is used for:
(A) Planning in large sequential decision problems
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Planning in large sequential decision problems
Q#37: Value function approximation helps with:
(A) Large state spaces
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Large state spaces
Q#38: Reward discounting ensures:
(A) Future rewards are valued less than immediate rewards
(B) Future rewards ignored completely
(C) BFS only
(D) DFS only
Answer: (A) Future rewards are valued less than immediate rewards
Q#39: Temporal abstraction in RL uses:
(A) Options or macro-actions for long-term planning
(B) Single-step actions only
(C) BFS only
(D) DFS only
Answer: (A) Options or macro-actions for long-term planning
Q#40: Exploration strategies include:
(A) ε-greedy, softmax, and upper confidence bound
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) ε-greedy, softmax, and upper confidence bound
Q#41: Delayed reward problem occurs when:
(A) Actions affect rewards received much later
(B) Immediate reward only
(C) BFS only
(D) DFS only
Answer: (A) Actions affect rewards received much later
Q#42: RL is inspired by:
(A) Behavioral psychology and trial-and-error learning
(B) Supervised learning only
(C) BFS only
(D) DFS only
Answer: (A) Behavioral psychology and trial-and-error learning
Q#43: Temporal difference error (TD error) measures:
(A) Difference between predicted and actual reward plus estimated value of next state
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Difference between predicted and actual reward plus estimated value of next state
Q#44: Convergence of Q-learning is guaranteed if:
(A) All state-action pairs are visited infinitely often and learning rate decays appropriately
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) All state-action pairs are visited infinitely often and learning rate decays appropriately
Q#45: Exploration ensures:
(A) Avoiding suboptimal policies due to lack of knowledge
(B) Always exploiting best-known action
(C) BFS only
(D) DFS only
Answer: (A) Avoiding suboptimal policies due to lack of knowledge
Q#46: Function approximation in RL can use:
(A) Linear functions, neural networks, or decision trees
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Linear functions, neural networks, or decision trees
Q#47: Deep reinforcement learning combines:
(A) Neural networks with RL algorithms
(B) BFS only
(C) DFS only
(D) Decision trees only
Answer: (A) Neural networks with RL algorithms
Q#48: Model-free RL learns:
(A) From experience without knowing the environment dynamics
(B) BFS only
(C) DFS only
(D) From transition models only
Answer: (A) From experience without knowing the environment dynamics
Q#49: Model-based RL can:
(A) Plan ahead using learned or known models
(B) BFS only
(C) DFS only
(D) Random guesses
Answer: (A) Plan ahead using learned or known models
Q#50: The main goal of reinforcement learning is:
(A) Learn policies that maximize long-term expected reward through interaction
(B) Memorize actions only
(C) BFS only
(D) DFS only
Answer: (A) Learn policies that maximize long-term expected reward through interaction