Core Concepts of AI: Decision Making, Problem Solving and Knowledge Reasoning

Decision-Making Frameworks in AI

1. Theoretical & Architectural Paradigms

• Human-like vs. Rational AI: AI decision-making operates under two primary philosophical approaches — Human-like AI, which strives to think and act like humans using cognitive modeling and the Turing Test approach, and Rational AI, which focuses on optimal decision-making through ideal rational agent behavior.
• Core Theoretical Frameworks: AI incorporates classical choice frameworks such as rationality, bounded rationality (satisficing under limited resources), and intuition.
• Operational Frameworks: AI architectures rely on model-based and algorithmic models, Human-in-the-Loop (HITL) systems for expert oversight, agentic and rules-based logic, governance and ethical frameworks, as well as utility and decision theory for sequential decision problems.

2. Data-Supported Decision-Making Pipeline

A data-supported decision framework converts raw environmental inputs into strategic choices through a 5-step process:

1. Collecting Data: Ingesting raw data from databases, APIs, or physical sensors.
2. Data Preprocessing: Cleaning data to remove errors, inconsistencies, and noise.
3. Data Analysis: Analyzing structured data to identify underlying patterns and operational trends.
4. Predictive Modeling: Using machine learning algorithms to build models that forecast future scenarios.
5. Decision Making: Applying analytical insights and predictive outputs to drive automated or human decision choices.

3. Core Capabilities (How AI Enhances Decisions)

• Data Analysis & Pattern Recognition: Analyzes vast datasets to surface subtle correlations that traditional methods overlook.
• Prediction & Optimization: Forecasts future market trends using historical data and balances multi-objective constraints to find optimal solutions (e.g., logistics route planning).
• Automation: Executes repetitive operational tasks using pre-set rules to minimize human error.
• Risk Management & Personalization: Detects anomalies to prevent threats (e.g., cybersecurity breaches) and customizes recommendations based on individual user behavior.

4. Agent-Based Architecture (PEAS Framework)

Decision-making in autonomous agents is defined by their interaction with the environment:

• Agent Anatomy: An AI Agent perceives its environment through sensors, processes digital inputs via an agent program, and acts upon the environment using actuators.
• PEAS Model: Systems are designed using four key parameters:
• Performance Measure (evaluation criteria)
• Environment (context and surroundings)
• Actuators (action mechanisms)
• Sensors (data input devices)

peas image

• Examples: Physical agents like self-driving cars (using LiDAR, radar, and cameras to execute steering and braking decisions) or virtual agents like customer support chatbots (using text feeds to generate ticketing and response actions).

5. Opportunities & Challenges

• Opportunities: Provides high speed and scale across thousands of variables, superior pattern recognition, predictive foresight, and standardized operational consistency.
• Challenges: Faces risks from algorithmic bias in training data, the black box problem (lack of deep learning explainability), data privacy/security risks, and human over-reliance during unexpected edge-case crises.

Problem-Solving Strategies and Their Types

1. Core Concept & Search Formulation

• Definition: Problem-solving in AI is a systematic process where an intelligent agent analyzes its environment, defines a target goal, and determines the best sequence of actions to reach that objective.
• 4 Components of a Search Problem: A well-defined search problem consists of an initial state, available actions, a transition model, and a goal test, evaluated alongside a path cost function.
• State Space vs. Search Tree: The state space is the set of all real physical configurations of the environment, while a search tree is the mathematical tree structure generated dynamically as the search algorithm runs.
• Evaluation Metrics: Search algorithms are evaluated on completeness (guaranteeing a solution if one exists), time complexity, space complexity, and optimality (finding the lowest path cost solution).

2. Types of Problem-Solving Strategies

A. Search-Based Strategies

• Uninformed (Blind) Search: Explores the state space blindly without domain knowledge or clues about the distance to the goal. Standard examples include Breadth-First Search (BFS), Depth-First Search (DFS), and Uniform Cost Search (UCS).
• Informed (Heuristic) Search: Employs a heuristic function (h(n)) to estimate the cheapest path cost from node (n) to the goal, restricting node expansion strictly to promising directions. Standard examples include *A Search and Greedy Best-First Search**.

B. Adversarial Search (Game-Playing)

• Tailored for multi-agent competitive environments where an active opponent works to defeat the AI agent.
• Utilizes techniques like the Minimax algorithm and Alpha-Beta pruning to evaluate game trees and make optimal counter-moves.

C. Constraint Satisfaction Problems (CSPs)

• Focuses on finding states that satisfy a rigid set of rules across three core elements: variables (items to decide), domains (allowed values), and constraints (unary, binary, or higher-order rules).
• Solved using backtracking search (systematically trying assignments and reversing when rules are violated), forward-checking (conducting local consistency checks on remaining variable domains), and constraint propagation.

D. Optimization and Search Approaches

• Linear Programming (LP): Solves straight-line optimization problems under linear constraints where optimal solutions always lie at the corner points of the rule-defined shape; widely used in resource allocation and supply chain routing.
• Genetic Algorithms (GA): Modeled after biological evolution and natural selection; maintains a population of candidate solutions evaluated by a fitness score, generating new populations via crossover and mutation.
• Simulated Annealing (SA): Inspired by metallurgical annealing; transitions to neighboring solutions and occasionally accepts worse choices at higher “temperatures” to escape local minima traps before gradually “cooling down” to settle on an optimal result.

Knowledge Inference, Logical Agents, and Knowledge-Based Agents

1. Knowledge Inference

• Definition: Knowledge inference is the act of deriving logical conclusions, new facts, or predictions from stored data, rules, or a structured knowledge base. It serves as the cornerstone of decision-making, allowing systems to reason dynamically rather than merely memorizing facts.
• Core Concepts:
• Knowledge Base (KB): A structured repository where facts, rules, and relationships are stored.
• Inference Engine: The reasoning component that applies logical rules to the knowledge base to derive new information, make predictions, or suggest actions.
• Entailment : A condition where a conclusion is guaranteed to be true if the supporting premise statements are true.

2. Logical Agent

• Definition: A Logical Agent is an AI system that relies on formal logic (such as Propositional Logic or First-Order Logic) to represent knowledge and reason about its environment.
• Distinction: Unlike simple reflex agents that respond directly to immediate percepts, a logical agent maintains an internal representation of its world, tracks changes over time, and deduces hidden properties to make sound decisions.
• Core Pillars: Operates using a specialized Knowledge Base (containing formal logical sentences) and an Inference Engine (applying domain-independent logical rules to evaluate action safety and derive new sentences).
• Operational Cycle (TELL, ASK, ACT):

1. TELL: Converts sensor percepts into formal logical sentences, adds them to the KB, and updates time.
2. ASK: Queries the KB to determine which action logically follows from current goals and known facts.
3. ACT: Executes the chosen action in the environment and logs the execution back into the KB.

3. Knowledge-Based Agent (KBA)

• Definition: A Knowledge-Based Agent (KBA) maintains an internal state of explicit knowledge, perceives its environment, updates its KB with observations, and uses an inference engine to reason and act intelligently. It contrasts with purely data-driven machine learning models that lack explicit knowledge representations.
• Three Levels of a KBA:

1. Knowledge Level: Focuses on what the agent knows (facts, rules, and domain relationships) to achieve its goals.
2. Logical Level: Deals with the structure and organization of knowledge using formal logic (e.g., “IF-THEN” rules like “IF a patient has a fever and rash, THEN consider measles”) to ensure logical consistency and enable deduction.
3. Implementation Level: Covers the computational data structures, algorithms, and hardware required to execute functions (e.g., using Python algorithms and SQL databases).

• Examples & Use Cases: Medical diagnosis systems (evaluating symptoms against stored medical rules to suggest treatments), expert systems, automated customer support chatbots, recommendation engines, and robotics task planning.

Propositional Logic Knowledge Representation and Semantic Networks

1. Propositional Logic Knowledge Representation

• Definition: Propositional logic is the simplest knowledge representation language in artificial intelligence. It represents factual information using declarative statements called propositions that evaluate strictly to true or false.

• Core Components & Rules:

Propositions: Simple or complex declarative statements assigned a clear truth value (e.g., “The sky is blue” is true; “It is raining” is false).

Logical Connectives: Used to combine simple propositions into complex compound formulas:

• AND : True only if both propositions are true.
• OR : True if at least one proposition is true.
• NOT : Reverses the truth value of a proposition.
• IMPLIES : Represents conditional reasoning in an "IF condition THEN conclusion" format.
• IF AND ONLY IF ((\leftrightarrow)): True if both propositions share the exact same truth value.

Truth Tables & Expression Types:

• Truth Tables: Evaluate logical expressions by testing all possible input truth combinations.

Tautology: An expression that is always true regardless of input values .

Contradiction: An expression that is always false for all inputs .

Contingency: An expression that can be true or false depending on input values .

• Logical Reasoning Concepts:

Entailment : A semantic relationship where a conclusion logically follows from premises.

Inference: The syntactic process of deriving new conclusions using formal operational rules.

Soundness & Completeness: Soundness ensures an inference system derives only true entailments, while completeness ensures it can derive all true entailments.

2. Semantic Networks

• Definition: A semantic network is a visual, graph-based knowledge representation technique that models concepts, objects, and relationships as an interconnected web.
• Core Components:

Nodes: Represent individual concepts, objects, entities, or events (e.g., “Bird”, “Dog”, or “John”).

Edges / Links: Directed arrows showing connections between nodes using specific labels like is-a, has-part, or can-do (e.g., linking "Dog" to "Animal" via an is-a edge).

Inheritance: A structural property where sub-concepts automatically inherit the attributes of broader parent concepts higher in the graph hierarchy (e.g., if “Bird” has the property “can fly”, “Sparrow” automatically inherits the ability to fly).

Uncertain Knowledge: Representing and Reasoning with Uncertain Knowledge

1. Meaning & Causes of Uncertain Knowledge

• Definition: Uncertain knowledge occurs when AI systems operate in environments where information is incomplete, noisy, ambiguous, or dynamic, making strict binary true/false logic impossible.
• Core Causes:
• Incomplete Data: Missing information blocks full visibility into a problem state.
• Noisy Sensors: Physical sensors (e.g., cameras, microphones) capture flawed signals.
• Ambiguity: Inputs or rules can have multiple valid interpretations.
• Dynamic / Stochastic Environments: Surroundings shift constantly based on random or probabilistic rules.
• Real-World Applications: Used in autonomous vehicles (unpredictable traffic/weather), healthcare diagnostics (symptom variability), natural language processing (contextual ambiguity), and fraud detection (uncertain transaction patterns).

2. Representing Uncertain Knowledge

Instead of binary true/false rules, probabilistic AI quantifies uncertainty using probabilities (values between 0 and 1), confidence scores, and outcome ranges. Major representation frameworks include:

• Probabilistic Reasoning & Bayes’ Rule: Uses probability theory to measure likelihoods and update beliefs when new evidence arrives. Bayes’ Rule calculates posterior probabilities (e.g., estimating whether an email is spam based on the presence of specific keywords like “win”).
• Bayesian Networks / Belief Networks: Directed Acyclic Graphs (DAGs) where nodes represent random variables, directed edges represent conditional dependencies, and Conditional Probability Tables (CPTs) specify outcome probabilities based on parent nodes (e.g., the Burglar Alarm network).
• Fuzzy Logic: Handles vague or continuous inputs by assigning degrees of truth between 0 and 1 (e.g., assigning 0.7 to “It is warm outside”) rather than strict true/false states.
• Dempster-Shafer Theory (DST / Theory of Evidence): Extends classical probability by assigning mass functions (basic probability assignments) to sets of hypotheses rather than single outcomes. It evaluates bounds using Belief functions (minimum confidence) and Plausibility functions (maximum possible belief).

3. Reasoning with Uncertain Knowledge

• Belief Updating & Evidence: Continuously updates likelihood estimates as new observations arrive (e.g., dark clouds appearing increases the estimated probability of rain). Enables both predictive reasoning (cause to effect) and diagnostic reasoning (effect to cause).
• Markovian Models:
• Markov Models: Predict future states based solely on the current state (memoryless property).
• Hidden Markov Models (HMMs): Infer unobservable “hidden” states using observable statistical data.
• Markov Decision Processes (MDPs / POMDPs): Model sequential decision-making and reinforcement learning in fully or partially observable uncertain environments.
• Inference & Learning: Uses exact methods (variable elimination) or approximate methods (Markov Chain Monte Carlo / MCMC) for computation, while updating model parameters over time via maximum likelihood or Bayesian estimation.

Key Advantages

Probabilistic reasoning enhances AI by offering flexibility across domains, robustness against noisy/missing data, transparency through explicit confidence scores, scalability for big data, and effective decision support under uncertainty.

Thank you for reading!! 😊😊