MI Flashcards
What is Supervised Learning?
Supervised learning is a machine learning approach that trains on labeled data, where each example has a known correct output or label.
Describe the structure of a Decision Tree.
A Decision Tree has: internal nodes for feature tests, branches for outcomes, and leaves for final classifications.
How is a Decision Tree constructed?
- Start with all data at the root. 2. Select a feature that best splits the data (e.g., using Gini or Information Gain). 3. Partition data accordingly. 4. Recursively repeat until a stopping condition (max depth, min samples, etc.) is reached.
Give an example of a Decision Tree classification.
Example: Predicting laptop purchases. If the person is a student, check budget. If budget > $1000, predict purchase.
What is the k-Nearest Neighbor (k-NN) algorithm?
k-NN is an instance-based learning method that classifies (or regresses) a new point based on its closest k neighbors.
Describe the k-NN algorithm.
- Choose k (number of neighbors). 2. Compute distance (e.g., Euclidean) from the new point to all training points. 3. Select the k closest. 4. Combine results via majority vote (classification) or average (regression).
What are key evaluation metrics in supervised learning?
Common metrics include: Accuracy (correct/total), Precision & Recall (especially for imbalanced data), and ROC AUC (trade-off between TPR and FPR).
What is a Neural Network?
A Neural Network is composed of: an input layer (features), one or more hidden layers (transformation of inputs), and an output layer (final predictions).
How does backpropagation train a Neural Network?
- Forward pass to get predictions. 2. Compute loss by comparing to true label. 3. Use gradients (via backpropagation) to update weights. 4. Repeat until convergence.
Where are neural networks commonly used?
Neural networks are used in image recognition, speech processing, and self-driving cars.
What is Unsupervised Learning?
Unsupervised learning discovers patterns from unlabeled data, unlike supervised learning which relies on labeled examples.
What is Clustering?
Clustering groups similar data points based on similarity (e.g., distance).
How does k-Means clustering work?
- Pick the number of clusters k. 2. Randomly choose k initial centers. 3. Assign each point to the nearest center. 4. Update centers based on assignments. 5. Repeat until centers stabilize.
How does Expectation-Maximization clustering work?
It alternates: E-Step (assign probabilistic memberships to clusters) and M-Step (update cluster parameters based on those memberships).
Where is clustering used?
Clustering is used in customer segmentation, text classification, and medical diagnosis.
What is a Constraint Satisfaction Problem (CSP)?
A CSP involves finding values for a set of variables within their domains so that all constraints are satisfied.
What are the key components of a CSP?
- Variables (items to assign). 2. Domains (possible values). 3. Constraints (rules that restrict valid assignments).
How can CSPs be solved?
- Backtracking Search (systematically try assignments). 2. Constraint Propagation (reduce domains before/during search). 3. Local Search (iteratively refine a complete assignment).
What is Generalized Arc Consistency (GAC) in CSPs?
GAC ensures that for every constraint, each value in a variable’s domain is consistent with some value in the domains of other variables in that constraint.
How does the Generalized Arc Consistency (GAC) algorithm work?
- Initialize a queue with all constraints. 2. Make a variable’s domain consistent for each constraint. 3. If values are removed, re-check related constraints. 4. Repeat until no more changes.
Why is Generalized Arc Consistency (GAC) useful?
It prunes inconsistent values early, reducing the search space and preventing needless backtracking.
What is a Bayesian Network?
A Bayesian Network is a directed acyclic graph whose nodes represent variables and edges indicate probabilistic dependencies.
How does inference work in a Bayesian Network?
- Exact Inference (e.g., variable elimination). 2. Approximate Inference (e.g., Gibbs sampling, MCMC).
What are some applications of Bayesian Networks?
They are used for medical diagnosis, spam filtering, robotics, and decision support systems.
What is a search problem in AI?
A search problem defines an initial state, a goal state, and actions to move between states. The task is to find a path to the goal (often with minimal cost).
What are some types of search algorithms?
- Uninformed (BFS, DFS, UCS). 2. Informed (Greedy, A*). 3. Local (Hill Climbing, Simulated Annealing).
How does A* search work?
A* uses f(n) = g(n) + h(n), where g(n) is the path cost so far and h(n) is a heuristic estimating the cost to the goal.
Where are search algorithms used?
Search algorithms are used in pathfinding, AI planning, game AI, and robotics.
What is Planning in AI?
Planning is selecting and organizing actions to transform an initial state into a goal state.
What are the main components of a planning problem?
It has states, actions, an initial state, a goal state, and a plan (sequence of actions).
What is STRIPS in planning?
STRIPS is a formalism that represents actions with preconditions, add lists, and delete lists.
What does the Delete-Relaxation technique do?
It ignores negative (delete) effects of actions to simplify planning, often used for heuristic generation.
What is a Multi-Agent System (MAS)?
A MAS involves multiple autonomous agents that can interact cooperatively or competitively.
What is the Minimax algorithm used for?
Minimax is used in adversarial games to minimize potential losses assuming an optimal opposing player.
What is Alpha-Beta pruning in Minimax?
Alpha-Beta pruning cuts off branches in the Minimax tree that won’t affect the final choice, speeding up the search.
What is a Nash Equilibrium?
A strategy profile in which no player can benefit by unilaterally changing their strategy.
What is Planning under Uncertainty?
Planning in environments where actions have uncertain or probabilistic outcomes.
What is a Markov Decision Process (MDP)?
An MDP is defined by (S, A, T, R, γ), where T is a transition function, R a reward function, and γ a discount factor.
What is the Bellman Equation used for?
The Bellman Equation iteratively updates value functions to find an optimal policy in an MDP.
What is Reinforcement Learning (RL)?
RL is where an agent learns from trial-and-error in an environment, receiving rewards or penalties for actions.
What is Q-Learning in RL?
A model-free RL method updating Q(s,a) via: Q(s,a) ← Q(s,a) + α [r + γ * max Q(s’,a’) – Q(s,a)].
What is an example of Reinforcement Learning?
A self-driving car learns optimal driving strategies through repeated trials and feedback (rewards/penalties).
How do agents in Multi-Agent Systems make decisions?
They may use game theory, reinforcement learning, or rule-based systems depending on cooperation or competition.
What is the role of a utility function in AI?
It represents an agent’s preferences, guiding action selection to maximize expected utility.
What is an example of using Nash Equilibrium in real life?
If one chain lowers its prices while the other keeps prices high, the lower‐priced chain will attract more customers and boost its profits (both do it == bad)
How does Value Iteration work in MDPs?
Value Iteration applies the Bellman update repeatedly to each state until values converge.
What is Policy Iteration in MDPs?
It alternates between policy evaluation (computing state values for a policy) and policy improvement until optimal.
What is the Exploration-Exploitation trade-off in RL?
Balancing trying new actions (exploration) and exploiting known high-reward actions to maximize long-term rewards.
What is a common application of MDPs?
Robotics, where a robot must plan and act under uncertain transitions.
What are some real-world applications of Planning?
Planning is used in robotics, autonomous driving, logistics, and scheduling tasks.
How does STRIPS represent actions?
STRIPS actions are defined by preconditions, add effects (what becomes true), and delete effects (what becomes false).
What is a heuristic function in AI planning?
A function estimating the remaining cost to reach a goal, used to guide search algorithms like A*.
What are some real-world applications of CSPs?
(Scheduling problems, Sudoku, Map Coloring, and Timetabling.)
What are the main techniques to solve CSPs?
(1. Backtracking Search, 2. Constraint Propagation, 3. Generalized Arc Consistency (GAC), 4. Local Search.)
How does Generalized Arc Consistency (GAC) work?
(GAC ensures that for every constraint, each value in a variable’s domain has a consistent corresponding value in other domains. It extends Arc Consistency to higher-order constraints.)
What are the key properties of a Bayesian Network?
(1. Models uncertainty using probability, 2. Encodes conditional dependencies, 3. Uses a Directed Acyclic Graph (DAG).)
What is conditional independence in Bayesian Networks?
(A variable is independent of its non-descendants given its parents.)
What is an example of a real-world search problem?
(Pathfinding in Google Maps, solving puzzles like the 15-Puzzle, or finding bugs in software.)
What are some common uninformed search strategies?
(Breadth-First Search (BFS), Depth-First Search (DFS), Uniform Cost Search (UCS).)
