Artificial Intelligence and Neural Networks

1

Artificial Intelligence (AI) is the field concerned with building machines that can perform tasks normally requiring human intelligence — reasoning, learning, perception, and problem solving.

2

AI has four perspectives: thinking humanly (cognitive modeling), thinking rationally (logic-based), acting humanly (the Turing Test approach), and acting rationally (the rational agent approach).

3

The term "Artificial Intelligence" was coined at the 1956 Dartmouth Conference, considered the birth of AI as a field.

4

AI's history includes periods of rapid progress, "AI winters" (reduced funding/interest), and a modern resurgence driven by machine learning and deep learning.

5

Applications of AI include robotics, natural language processing, computer/machine vision, expert systems, game playing, autonomous vehicles, and recommendation systems.

6

An agent is anything that perceives its environment through sensors and acts upon that environment through actuators; the agent function maps a percept sequence to an action.

7

Structure of an agent = architecture (the underlying hardware/platform) + agent program (the software implementing the agent function).

8

Properties of intelligent agents: autonomy (operates without direct human intervention), reactivity (responds to environment changes), proactiveness (takes initiative to achieve goals), and social ability (interacts with other agents/humans).

9

A simple reflex agent selects actions based only on the current percept using condition-action rules; it has no memory of the past.

10

A model-based reflex agent maintains an internal model/state of the world to handle partial observability, in addition to condition-action rules.

11

A goal-based agent chooses actions that will achieve a defined goal, considering the future consequences of actions — unlike a simple reflex agent, which reacts only to the current percept.

12

A utility-based agent chooses actions that maximize a utility function (a measure of "happiness"/performance), useful when multiple goals conflict.

13

Environment types (paired contrasts): Deterministic (next state fully determined by current state and action) vs Stochastic (an element of randomness/uncertainty).

14

Further environment contrasts: Static (unchanged while deliberating) vs Dynamic (can change during deliberation); Fully observable (sensors give complete state info) vs Partially/semi-observable (incomplete info); Single-agent vs Multi-agent (multiple agents interact, cooperating or competing).

1

A problem is formulated using: initial state, a set of possible actions, a transition model (result of an action), a goal test, and a path cost function.

2

A well-defined problem has all of the above clearly specified, allowing systematic search for a solution.

3

A Constraint Satisfaction Problem (CSP) is defined by a set of variables, their domains (possible values), and a set of constraints that must be satisfied (e.g., map coloring, Sudoku, the N-queens problem).

4

Breadth-First Search (BFS) explores the search tree level by level; complete and optimal for uniform step costs, but requires a lot of memory.

5

Depth-First Search (DFS) explores as far as possible along each branch before backtracking; uses little memory, but is not guaranteed complete or optimal.

6

Depth-Limited Search (DLS) is DFS restricted to a pre-set maximum depth, avoiding infinite paths.

7

Iterative Deepening Search (IDS) repeats DLS with gradually increasing depth limits; combines BFS's completeness/optimality with DFS's low memory usage.

8

Bidirectional Search runs two simultaneous searches — forward from the start and backward from the goal — meeting in the middle to reduce search effort.

9

Greedy Best-First Search expands the node that appears closest to the goal, using only the heuristic h(n); fast, but not guaranteed optimal.

10

**A\* Search** expands the node with the lowest f(n) = g(n) + h(n) (cost so far + estimated cost to goal); guaranteed optimal if the heuristic is admissible (never overestimates).

11

Hill Climbing is a local search that always moves to the neighboring state with a better value; simple, but can get stuck at local maxima, plateaus, or ridges.

12

Simulated Annealing is similar to hill climbing but occasionally accepts a worse move (based on a decreasing "temperature" schedule) to escape local optima.

13

Adversarial search is used in two-player, zero-sum games where one player's gain is the other's loss.

14

The Minimax algorithm computes the optimal move by assuming the maximizing player tries to maximize the score while the minimizing player (opponent) tries to minimize it; Alpha-Beta pruning optimizes minimax by pruning branches that cannot possibly influence the final decision, without changing the result.

1

Knowledge representation (KR) is the field concerned with representing information about the world in a form a computer system can use to reason and solve problems.

2

KR approaches include: logical representation (formal logic), semantic network representation, frame-based representation, and procedural representation.

3

Issues in KR include: choosing important attributes to represent, representing relationships among objects, choosing the right level of granularity, representing sets of objects, and finding the right structural framework.

4

Semantic networks represent knowledge as a graph with nodes (objects/concepts) connected by labeled edges representing relationships (e.g., "is-a", "has-part").

5

Frames are data structures for representing stereotyped situations or objects, organized into named slots holding fillers (values or default values).

6

Propositional Logic (PL) uses atomic propositions combined with logical connectives: AND (∧), OR (∨), NOT (¬), IMPLIES (→), and IFF (↔).

7

A well-formed formula (WFF) is a syntactically valid logical expression; a tautology is a formula that is always true regardless of the truth values of its components; validity means a formula is true under every interpretation.

8

Resolution is an inference rule (used in refutation-based proofs): it converts statements to clausal (CNF) form and repeatedly resolves pairs of clauses to derive a contradiction, proving a conclusion follows from the premises.

9

First-Order Predicate Logic (FOPL) extends propositional logic with objects, predicates, and quantifiers: the universal quantifier (∀, "for all") and the existential quantifier (∃, "there exists").

10

Rules of inference (e.g., Modus Ponens, Universal Instantiation) allow new facts to be derived from known facts; unification finds a substitution of variables that makes two logical expressions identical, essential for applying inference rules and resolution in FOPL.

11

Bayes' Rule: P(A|B) = P(B|A)·P(A) / P(B) — used to update the probability of a hypothesis A given new evidence B.

12

A Bayesian Network is a directed acyclic graph (DAG) where nodes represent random variables and edges represent probabilistic dependencies between them.

13

Reasoning in belief networks involves computing the posterior probability of certain variables given observed evidence about others (probabilistic inference).

1

An expert system is a computer program designed to emulate the decision-making ability of a human expert within a specific, narrow domain.

2

Expert system components: Knowledge base (stores domain facts and rules), Inference engine (applies logical rules to derive new facts/conclusions), User interface (allows interaction), Explanation facility (explains the reasoning behind a conclusion), and Knowledge acquisition module (supports adding/updating knowledge).

3

Knowledge acquisition is the process of extracting, structuring, and organizing knowledge from human experts (or other sources) into the expert system's knowledge base.

4

Declarative knowledge ("knowing what" — facts and relationships) is distinguished from procedural knowledge ("knowing how" — encoded as steps/procedures).

5

Developing an expert system typically involves a knowledge engineer working closely with domain experts to elicit and encode their expertise.

6

NLP terminology includes: syntax (grammatical structure), semantics (meaning), pragmatics (context-dependent meaning), morphology (word structure), and phonology (sound structure).

7

Natural Language Understanding (NLU) interprets/extracts meaning from human language input; Natural Language Generation (NLG) produces human language output from structured data.

8

Steps of NLP: lexical analysis (tokenizing words) → syntactic analysis/parsing (grammatical structure) → semantic analysis (meaning) → discourse integration (context across sentences) → pragmatic analysis (real-world intended meaning).

9

Applications of NLP: machine translation, chatbots/virtual assistants, sentiment analysis, speech recognition, and text summarization.

10

NLP challenges: lexical/syntactic/semantic ambiguity, understanding context, handling idioms and sarcasm, and resolving pronoun references.

11

Machine vision enables machines to interpret and understand visual information from the world (images/video).

12

Stages of machine vision: image acquisition → preprocessing (noise removal, enhancement) → segmentation (isolating regions of interest) → feature extraction → recognition/interpretation.

13

Robotics combines AI, mechanical engineering, and electronics to build machines (robots) that can sense their environment, plan actions, and physically act upon the world.

1

Machine learning enables systems to improve their performance from experience/data without being explicitly programmed for every scenario.

2

Supervised learning learns a mapping from inputs to outputs using labeled training data (input-output pairs); used for classification and regression.

3

Unsupervised learning finds patterns/structure in unlabeled data; used for clustering and association.

4

Reinforcement learning: an agent learns by interacting with an environment, receiving rewards or penalties for its actions, aiming to maximize cumulative reward.

5

Inductive learning generalizes rules from specific examples; a decision tree is a common inductive learning model that splits data based on attribute tests (often chosen using information gain/entropy) to reach a classification.

6

Naive Bayes is a statistical-based, probabilistic classifier based on Bayes' theorem that assumes features are conditionally independent given the class.

7

Fuzzy learning/logic handles reasoning that is approximate rather than fixed/exact, using fuzzy sets where an element's membership can range continuously between 0 and 1 (rather than strictly 0 or 1).

8

A Fuzzy Inference System (FIS) maps inputs to outputs through three stages: fuzzification (converting crisp inputs to fuzzy values), rule evaluation (applying fuzzy if-then rules), and defuzzification (converting the fuzzy result back to a crisp output).

9

Fuzzy inference methods include the Mamdani method (widely used, intuitive, fuzzy output) and the Sugeno method (computationally efficient, often used in control systems).

10

A Genetic Algorithm (GA) is an optimization technique inspired by natural selection and evolution, working with a population of candidate solutions (chromosomes).

11

GA operators: selection (choosing fitter individuals to reproduce), crossover (combining parts of two parent chromosomes), and mutation (randomly altering parts of a chromosome to maintain diversity).

12

Encoding represents candidate solutions as chromosomes, e.g., binary strings or real-valued vectors.

13

Selection algorithms include roulette wheel selection (probability proportional to fitness), tournament selection, and rank-based selection.

14

The fitness function evaluates how good a candidate solution is at solving the problem; GA parameters include population size, crossover rate, mutation rate, and the number of generations to run.

1

A biological neuron has dendrites (receive signals), a cell body (processes signals), an axon (transmits output), and synapses (connections to other neurons).

2

An Artificial Neural Network (ANN) models this using artificial neurons: inputs (with associated weights), a summation function (weighted sum), and an activation function that produces the output.

3

The McCulloch-Pitts neuron was the earliest mathematical model of a neuron — a simple binary threshold unit that fires (outputs 1) if the weighted sum of inputs exceeds a threshold.

4

Step (threshold) activation outputs a fixed value (e.g., 0 or 1) based on whether the input exceeds a threshold; Sigmoid activation is an S-shaped curve with output between 0 and 1, smooth and differentiable.

5

Tanh activation is an S-shaped curve with output between -1 and 1 (a rescaled sigmoid); ReLU outputs the input directly if positive, else 0 (max(0, x)), popular in deep learning for its simplicity and efficiency.

6

Neural network architectures include feedforward networks (signals flow one direction, input to output, no loops) and recurrent networks (contain feedback loops, allowing memory of previous states).

7

The Perceptron is the simplest neural network model — a single-layer linear classifier that can only correctly classify linearly separable data.

8

The learning rate controls the size of the weight adjustment made during each training step; Gradient descent is an optimization algorithm that iteratively adjusts weights in the direction that most reduces the error (the negative gradient of the error function).

9

The Delta rule (Widrow-Hoff rule) updates weights in proportion to the error between the desired output and the actual output.

10

Hebbian learning follows the principle "neurons that fire together, wire together" — the connection weight between two neurons increases when both are activated simultaneously.

11

Adaline (Adaptive Linear Neuron) is similar to the perceptron but uses a linear activation function during training and is trained using the delta rule (least mean squares).

12

A Multilayer Perceptron (MLP) has one or more hidden layers between the input and output layers, allowing it to solve non-linearly separable problems that a single-layer perceptron cannot.

13

The Backpropagation algorithm trains multilayer networks by propagating the output error backward through the network, layer by layer, and updating each weight using gradient descent to minimize the overall error.

14

A Hopfield network is a type of recurrent neural network used as an associative (content-addressable) memory, capable of recalling a complete stored pattern from a partial or noisy input.