Author: sevanand yadav

Prof. Indumathi

Introduction

Prerequisites

1. Basic probability & statistics.
2. Propositional/Boolean logic, laws, and rules.
3. Fundamental algorithms and data structure concepts.
4. Coding proficiency in Python.

Scope : Covers fondation for agentic AI.

Module 1’s agenda – PEAS Framework: The PEAS framework (Performance Measure, Environment, Actuators, Sensors) .

Crux: How to represent knowledge using probabilistic way.

Index

• Terminology

Introduction to Core AI Terminology (via Navigation Example)

AI terminology using a self-autonomous car / traveling agent toy example:

1. Agent / Intelligent Agent: An entity capable of perceiving its environment and taking actions that mimic human reasoning or behavior.
2. Environment: The external world or context in which the agent operates (e.g., the map/locality).
3. Exploration: The process by which an agent traverses unknown paths when no prior map data is available.
4. Exploitation: Using the shortest known route already stored in its Knowledge Base to travel from the Initial State to the Goal State as quickly as possible.
5. Fact – Learnings from exploration.
6. Knowledge Base (KB): The agent’s memory where explored routes, map layouts, and learned facts are stored for future decision-making.
7. Sensors & Percepts:
   • Sensors: Hardware/software tools (cameras, LIDAR, etc.) used to observe the environment.
   • Percept: The specific input or sensory observation received from the environment at any given time.
8. Actuators & Actions:
   • Actuators: Mechanisms enabling the agent to interact with the environment.
   • Actions: Decisions executed by the agent (e.g., move forward, turn left/right, apply brakes).
9. Initial State & Transition Model:
   • Initial State: The starting point/condition of the problem (e.g., the Source location).
   • Transition Model: The logical framework defining how taking a specific action from a given state shifts the agent into a new resultant state (e.g., moving right from Source leads to Junction 1).

The Core Problem: Why Balance is Necessary

• Pure Exploitation: If an agent only exploits what it knows early on, it acts greedily. It will get stuck in a local optimum because it never attempts actions that could lead to much higher rewards overall.
• Pure Exploration: If an agent spends all its time exploring, it wastes compute power, time, and resources randomly sampling paths without ever taking advantage of the high-reward paths it already found.

Reference:

⁠Artificial Intelligence: A Modern Approach (4th Edition) by Russell & Norvig.

Index

1. Inteliggence
2. 4 perspective of AI
3. State space representation
4. Environment

What is Intelligence?

• Definition: The capacity to perceive environmental data, reason from available facts, plan/take decisions, and learn/adapt to patterns.
• Key Observations:
  • Are humans always intelligent? No—humans make mistakes and get emotional.
  • Can animals be intelligent? Yes (e.g., dogs tracking scent, dolphins communicating, cats sensing earthquakes early).
  • Can machines be intelligent? Yes (e.g., Chess AI, Medical AI, Autonomous driving).
  • Takeaway: Intelligence is not exclusive to human beings.
  • Human vs. AI Learning: Humans learn natively by observing and imitating; AI systems rely on structured data feeding (50%–80%+ initial training datasets) before adapting.

The Four Perspectives of AI (Russell & Norvig Framework)

State-Space Representation :

In Data Structures & Algorithms (DSA) and System Design interviews, candidates are often evaluated on how compactly they represent system states.

As demonstrated with the Vacuum Cleaner World (Location, Status_A, Status_B), storing raw image frames or bloated objects causes state-space explosion and memory overhead.

Real-World Application

• Game AI & Simulators: Games like Chess or Go use compact bitboards (bitmasks) to represent pieces on a board rather than complex 2D arrays, speeding up state evaluation by orders of magnitude.
• Navigation & Routing (Google Maps, Uber): Road networks are compressed into state graphs where states are intersections and edges are road segments with dynamic weight (traffic latency).

Task Environment Classification (PEAS Framework)

• PEAS Profile for Automated Taxi Driver:
  • P (Performance Measure): Safe, fast, legal, comfortable trip; profit maximization.
  • E (Environment): Roads, traffic, pedestrians, weather.
  • A (Actuators): Steering, accelerator, brake, signal, horn.
  • S (Sensors): Cameras, LIDAR, speedometer, GPS.

Environment Properties:

1. Fully Observable vs. Partially Observable:
   1. Whether sensory input provides complete state visibility or limited/restricted visibility (e.g., hidden rooms or obstructed views).
   2. Fully observable environments provide complete visibility and a complete knowledge base, whereas partially observable ones have incomplete information and uncertainty.
2. Deterministic vs. Stochastic: Whether executing action A in state S always yields the exact same outcome $S’$, or introduces uncertainty.
   1. Deterministic environments yield the same output for a given input-action pair in a certain environment, while stochastic environments deal with uncertain events where outcomes vary.
3. Episodic vs. Sequential: Whether current actions affect future decisions (sequential) or if each perceptual episode is independent (episodic).
   1. In sequential environments, previous decisions impact current decisions; in episodic environments, current decisions are independent of prior actions.
4. Static vs. Dynamic
   1. Static: The environment remains unchanged while the agent is acting/thinking.
   2. Dynamic: The environment changes continuously or updates alongside agent actions.
   3. Semi-Dynamic: The physical environment does not change with time, but the agent’s performance score or evaluation metrics change over time.
5. Discrete vs. Continuous
   1. Environments are discrete if the states and input variables are finite and countable (e.g., grid vacuum cleaner, Rubik’s cube), and continuous if states are infinite/uncountable (e.g., autonomous driving).

Index

1. Agent architecture
2. Problem solving agent
3. Search

Note : important as agent architecture discussed.

Agent Architectures & Functions

Core Agent Architectures

1. Table-Driven / Table Lookup Agent: Maps percept sequences directly to actions; conceptually simple but suffers from massive lookup table sizes and slow generation.
   1. Suffers from massive lookup table sizes and unscalable learning overhead.
2. Simple Reflex Agent: A rule-based (condition-action) system acting only on the current percept in fully observable environments without considering future consequences.
   1. based only on the current percep
3. Model-Based Reflex Agent: Maintains internal state and transition models to track how the world evolves and handle partially observable/uncertain environments.
4. Goal-Based Agent: Uses search and planning techniques to evaluate whether potential actions will move the agent closer to a predefined goal state.
5. Utility-Based Agent: Extends goal-based agents by using a utility function (e.g., measuring cost, time, distance, safety) to select the best or most optimal path among multiple goal-reaching alternatives.
6. Learning Agent: Composed of four components:
   • Performance Element: Selects external actions based on current percepts.
   • Critic: Evaluates the agent’s behavior against a performance standard and provides positive/negative feedback.
   • Learning Element: Modifies and improves future performance based on critic feedback.
   • Problem Generator: Suggests suboptimal/unseen actions to encourage exploration of new states.

Problem Solving Agents & Problem Formulation

• 5 Core Problem Components:
  1. Initial State: The agent’s starting condition.
  2. Actions: Legal moves available to the agent.
  3. Transition Model / Successor Function: What state results from taking a given action in a specific state ($Result(s, a)$).
  4. Goal Test: Checks whether a given state matches the destination/objective configuration.
  5. Path Cost ($g(n)$): Cumulative cost or distance associated with transitions.
• Worked Problem Formulation Examples:
  • Romania map navigation (e.g., Arad to Bucharest).
  • Vacuum-cleaner world.
  • 8-Puzzle problem (state representation, $9!/2$ reachable state space, swapping with empty tile).

Search Concepts & Breadth-First Search (BFS)

• Key Terminology:
  • State vs. Node: A state represents the environment configuration; a node is a data structure in the search tree containing state, parent, actions, and path cost.
  • Frontier / Open List: Queue/structure holding nodes discovered but yet to be expanded.
• Uninformed vs. Informed Search:
  • Uninformed (Blind): Uses only basic problem formulation details without knowing how close a non-goal state is to the goal.
  • Informed (Heuristic): Uses domain knowledge via a heuristic function $h(n)$ estimating the remaining cost from state $n$ to the goal.
• Breadth-First Search (BFS) Mechanics & Complexity:
  • Explores the search space level-by-level (breadth-wise) using a FIFO queue.
  • Goal test is performed on node generation.
  • Time Complexity: $O(b^{d+1})$, where $b$ is the maximum branching factor and $d$ is the depth of the shallowest goal.
  • Space Complexity: $O(b^{d+1})$ / $O(b^d)$ to store all frontier and generated nodes in memory.

Topics Discussed

1. Probabilistic Representation & Reasoning
   1. Probability concept for uncertainty
2. Bayesian Network
   1. CPT
   2. Joint probability
   3. Conditional Probability
   4. Inference -4 Types depending on position of Q-(query)
3. Variable elimination
4. D-contentedness vs D-Separation -if they are conditionally independent.
5. refer ppt for handwritten solution at end of ppt – 8-9 marks