Core-2

Q1. Define intelligent agents. Explain their architecture with suitable examples. Compare typical and rational agents. (20 Marks)

1. Introduction to Intelligent Agents

In the field of Artificial Intelligence (AI), an "Agent" is anything that can view and understand its surrounding environment through "Sensors" and then act upon that environment using "Actuators." An Intelligent Agent is an advanced version of this. It does not just act randomly; it processes the information, thinks, makes decisions, and acts in a way that helps it achieve a specific goal.

Examples of agents include:

• Human Agent: Has eyes and ears as sensors, and hands and legs as actuators.
• Robotic Agent: Has cameras and infrared range finders as sensors, and motors as actuators.
• Software Agent: Has keystrokes and file contents as sensors, and screen displays or network packets as actuators.

2. Architecture of Intelligent Agents

The architecture of an intelligent agent is the physical or software machinery that the agent runs on. The basic formula is:

Agent = Architecture + Agent Program

• Architecture: The computing device, PC, or robotic hardware with physical sensors and actuators.
• Agent Program: The logic, software, or algorithm that runs inside the architecture to make decisions.

Types of Agent Architectures:

• Simple Reflex Agents: These are the most basic agents. They operate only on the current situation (percept) and ignore the past history. They use "Condition-Action rules" (If-Then rules).
  • Example: A smart thermostat. IF the temperature is above 30 degrees, THEN turn on the AC.
• Model-Based Reflex Agents: These agents maintain an internal state or memory. They keep track of the world that they cannot see right now.
  • Example: A self-driving car. If a pedestrian walks behind a large bus, the car cannot see them anymore, but its "internal model" remembers the pedestrian is there and applies the brakes.
• Goal-Based Agents: Knowing the current state is not always enough; the agent needs goal information. These agents plan their actions to reach a specific destination.
  • Example: A pathfinding robot in a maze calculating multiple different paths to figure out which one leads to the exit.
• Utility-Based Agents: Sometimes there are multiple ways to reach a goal, but some ways are better, faster, or safer. "Utility" means how happy or successful the agent is.
  • Example: A GPS navigation app (like Google Maps) picking the route with the least traffic and shortest time.

3. Typical Agents vs. Rational Agents

• Typical Agent: A typical agent simply follows the rules it was programmed with. It does not care if the outcome is the best possible outcome. For example, a robotic vacuum cleaner that just bounces off walls randomly.
• Rational Agent: A rational agent is designed to be optimal. A rational agent always chooses the action that is expected to maximize its performance measure based on the information it has collected so far.
  • Comparison: A rational vacuum agent will scan the room, map the dirty areas, calculate the shortest path to clean everything, and return to its charging dock before the battery dies.

Q2. Explain the architecture of software agents with a diagram. Describe negotiation and bargaining strategies among agents with examples. How does argumentation improve communication? (20 Marks)

1. Architecture of Software Agents

Unlike physical robots, Software Agents (often called "Softbots") live entirely inside computer systems, networks, or the internet. Instead of cameras, they use APIs, databases, and user inputs to sense the world. Instead of motors, they use system commands, emails, and database updates to act.

Components of Software Architecture:

• Communication Interface: This handles all incoming and outgoing data. It reads web pages, receives messages from other agents, and outputs results.
• Reasoning/Decision Engine: This is the brain of the software agent. It takes the input data, applies logic, solves problems, and decides what command to run next.
• Knowledge Base: This is the memory of the agent. It stores rules, historical data, user preferences, and facts about the environment.

2. Negotiation and Bargaining Strategies

In a system where multiple agents work together (Multi-Agent System), they often need to share resources. To do this without fighting, they use negotiation.

• Competitive Negotiation (Zero-Sum): In this strategy, agents have opposite goals. If Agent A wins, Agent B loses.
  • Example: Stock market trading agents. Agent A wants to buy shares at the lowest price, and Agent B wants to sell at the highest price.
• Cooperative Negotiation (Win-Win): In this strategy, agents work together to maximize the benefit for the whole system. They are willing to make compromises.
  • Example: Smart traffic light agents cooperatively keeping lights red a bit longer to let a jam clear, benefiting the whole city's traffic flow.

3. Argumentation in Communication

Normally, agents communicate using simple data commands (e.g., "Accept Proposal" or "Reject Proposal"). Argumentation makes this communication much smarter. It means that when an agent says "No," it also provides a reason or justification.

• How it improves communication: If Agent A asks Agent B to schedule a meeting at 2:00 PM, a normal agent would just reply "Error." With argumentation, Agent B replies: "I reject 2:00 PM because the server is down."
• Because Agent A now knows the reason, it can improve its next request to a different time. This prevents system deadlocks.

Q3(a). Differentiate between informed and uninformed search strategies with examples. What are constraint satisfaction problems? Explain with an example. (12 Marks)

1. Informed vs. Uninformed Search Strategies

• Uninformed Search (Blind Search):
  • Definition: This search strategy has no extra information about the problem other than the basic rules. It just blindly searches every possible path step-by-step.
  • Characteristics: Very slow, takes up a lot of memory, but guarantees a solution if one exists.
  • Examples: Breadth-First Search (BFS) and Depth-First Search (DFS).
• Informed Search (Heuristic Search):
  • Definition: This strategy uses domain knowledge or a "Heuristic" (a smart guess) to estimate how close it is to the goal.
  • Characteristics: Much faster and highly efficient for complex, real-world problems.
  • Examples: A* (A-Star) Search. When finding a route on a map, A* uses the straight-line distance to the destination as a hint.

2. Constraint Satisfaction Problems (CSPs)

Instead of looking for a path, the agent is looking for a "state" that follows a strict set of rules. A CSP consists of three things:

1. Variables: The items that need to be assigned a value.
2. Domains: The list of possible values that can be given to the variables.
3. Constraints: The rules that must not be broken.

• Example: The Map Coloring Problem
  • Variables: Different states on a map.
  • Domain: Three crayons (Red, Green, and Blue).
  • Constraint: No two states that share a border can have the same color.
  • The AI must assign a color to every state without breaking this rule.

Q3(b). Write short note on alpha-beta pruning in game playing. (8 Marks)

Introduction to Alpha-Beta Pruning

In AI game playing, agents use an algorithm called Minimax. The problem is that in games like Chess, there are millions of possible moves. The computer cannot calculate all of them in time. Alpha-Beta Pruning is an optimization technique used to make the Minimax algorithm faster.

How it Works:

• Alpha (α): Represents the best (highest) value the AI has found so far.
• Beta (β): Represents the best (lowest) value the opponent has found so far.

While checking future moves, if the AI finds a path that is clearly worse than a move it has already discovered (when Alpha ≥ Beta), it stops checking that path completely.

Conclusion:

Cutting off these useless branches is called "Pruning." Because the AI no longer wastes time calculating bad moves, it can look much deeper into the game tree, making the game-playing agent significantly stronger.

Q4. How does backtracking search improve performance in constraint-based problems? Analyze the effectiveness of stochastic games in representing real-world uncertain environments. (20 Marks)

1. Backtracking Search in Constraint-Based Problems

A simple AI might try to assign all values randomly and check at the very end if rules were broken (Generate-and-Test), which is very slow.

• Step-by-Step Checking: Backtracking assigns a value to one variable and immediately checks if a rule is broken.
• Stepping Back (Backtracking): If the rule is broken, the AI immediately stops, deletes that value, and tries the next available value or goes back to the previous variable.
• Huge Time Savings: By catching the mistake early, the AI avoids searching millions of future combinations that were guaranteed to fail anyway.

2. Stochastic Games and Uncertain Environments

Traditional games like Chess are "Deterministic." Stochastic games include elements of chance, luck, and randomness (like rolling dice in Ludo or drawing cards in Poker).

• Effectiveness in the Real World: The real world is highly unpredictable. If you are programming a Self-Driving Car or a Stock Market AI, you cannot predict exactly what will happen next.
• Stochastic game algorithms (like Expectiminimax and Markov Decision Processes) are highly effective here. Instead of planning for one exact future, they calculate the probability of different futures, allowing the AI to make the safest move even if things go wrong.

Q5(a). Discuss the major challenges faced in NLP and how modern approaches address them. (10 Marks)

1. Major Challenges in Natural Language Processing (NLP)

• Ambiguity (Double Meanings): Human words have multiple meanings. For example, "I am going to the bank." (River bank or financial bank?).
• Syntax and Grammar Complexity: Every language has different grammar rules, exceptions, and sentence structures.
• Sarcasm, Slang, and Idioms: Computers struggle to understand literal versus sarcastic meanings.
• Context: Human conversations rely on memory. Computers struggle to link pronouns ("He", "It") to proper nouns over long paragraphs.

2. How Modern Approaches Address Them

In the past, programmers used Rule-based NLP, which failed because language is too complex. Today, we use Deep Learning and Large Language Models (LLMs) like Transformers, BERT, and GPT.

• Self-Attention Mechanism: Modern AI looks at the entire sentence at once, calculating mathematical relationships between all words to figure out context.
• Vast Training Data: Instead of hand-written rules, models are trained on the entire internet, learning slang, idioms, and grammar naturally.

Q5(b). Illustrate the application of finite state automata in lexical analysis. (10 Marks)

1. What is Lexical Analysis?

When an AI reads human text, Lexical Analysis chops the giant string of letters into meaningful chunks called "Tokens" (e.g., breaking a sentence into Pronoun, Verb, and Number).

2. Application of Finite State Automata (FSA)

To do this chopping quickly, the system uses a Finite State Automata (FSA), an abstract mathematical machine that moves through different "states" until it recognizes a valid pattern.

How it works (Illustration): Suppose the AI needs to identify a valid Number.

• Start State: The machine is waiting.
• Input '5': The machine sees a digit and moves to the "Number State."
• Input 'A': If it sees a letter next, it breaks the rule and goes to an "Error State."
• Accepting State: If it only reads numbers and then a space, it stops at an "Accepting State" and tags it as a Number Token.

Q6(a). Explain smoothing techniques and their significance in probabilistic models. (10 Marks)

1. The Problem in Probabilistic Models

In text generation, models calculate the probability of the next word based on training data. The issue is the Zero-Frequency Problem: If the AI encounters a word it has never seen, the frequency is 0. Since probabilities are multiplied, a single 0 turns the entire sentence probability to 0, crashing the system.

2. Smoothing Techniques and Significance

Smoothing techniques are mathematical tricks that "smooth" out the data by taking a little bit of probability from common words and giving it to unseen words.

• Laplace Smoothing (Add-1 Smoothing): Simply adds the number 1 to the count of every single word in the dictionary, ensuring no word has a 0 count.
• Good-Turing Smoothing: Looks at the probability of words that appeared exactly once, and uses that math to estimate the probability of words that have not appeared yet.

Significance: Smoothing is absolutely crucial to prevent speech recognition or language translators from breaking down when encountering new names, rare words, or typos.

Q6(b). Compare rule-based and statistical part-of-speech tagging. (10 Marks)

What is Part-of-Speech (POS) Tagging?

POS tagging is the process of labeling every word in a sentence with its correct grammar tag (Noun, Verb, Adjective, etc.).

1. Rule-Based POS Tagging

• How it works: Uses massive dictionaries and thousands of grammar rules manually written by human linguists.
• Pros: Highly accurate for formal, perfectly structured text.
• Cons: Very slow and expensive to create. Fails easily if someone uses poor grammar or internet slang.

2. Statistical POS Tagging

• How it works: Uses Machine Learning (like Hidden Markov Models) trained on a massive database of pre-tagged text to learn the mathematical probability of words acting as specific parts of speech.
• Pros: Very fast to develop. It can learn new languages automatically and handles spelling mistakes better.
• Cons: Requires a massive amount of pre-tagged training data to become accurate.

Q7(a). Discuss the evolution, vision and enabling technologies of IoT. (10 Marks)

1. Evolution of IoT

• Stage 1: Barcodes and RFID tags used to give objects digital identities.
• Stage 2: Smart devices (Phones/PCs) connected to the internet.
• Stage 3: Modern IoT, where everyday objects are embedded with tiny chips to communicate autonomously over the internet.

2. The Vision of IoT

The ultimate vision is a fully connected, smart world where the physical environment responds automatically to human needs (like your car telling your house to turn on the AC), improving efficiency and saving energy.

3. Enabling Technologies

• Miniaturized Sensors: Cheap, tiny sensors that require almost no electricity.
• Wireless Connectivity: Technologies like 5G, Wi-Fi 6, Zigbee, and Bluetooth Low Energy.
• Cloud Computing: Powerful servers that process the massive data sent by sensors.
• IPv6: Provides enough unique internet addresses for every single sensor on the planet.

Q7(b). How do M2M communication and IoT differ in context and architecture? (10 Marks)

1. Context Differences

• M2M (Machine-to-Machine): Older concept involving point-to-point communication between two specific machines for a specific task (e.g., an ATM securely communicating with a bank server).
• IoT: Broader concept. A massive network of devices sharing data globally across multiple platforms and apps.

2. Architecture Differences

• M2M Architecture: Hardware-centric, operates on a closed network (private data between two machines), and has a limited scope without relying on Cloud computing.
• IoT Architecture: Software and Data-centric, operates on open IP networks (internet), and completely relies on Cloud platforms (like AWS) for processing large amounts of data.

Q8(a). Discuss security and standardization issues in IoT with reference to real-life case studies. (10 Marks)

1. Security Issues in IoT

IoT devices are designed to be low-cost and low-power, meaning they lack processing power for heavy encryption, making them the weakest link in cybersecurity.

• Real-Life Case Study (The Mirai Botnet, 2016): Hackers infected millions of cheap IoT security cameras that used default factory passwords. They then used these cameras to execute a massive DDoS attack, crashing major websites like Twitter and Netflix.

2. Standardization Issues

The IoT market is highly fragmented because major companies built their own separate communication protocols.

• The Problem: Lack of interoperability (a smart lock from Amazon might not talk to a smart speaker from Apple).
• The Solution: The industry is creating unified standards, like the "Matter" protocol, to ensure devices across brands can understand each other.

Q8(b). Explore the challenges of integrating IoT in smart cities, including data handling and cost-efficiency. (10 Marks)

1. Data Handling Challenges (Big Data)

• Volume and Velocity: Smart cities generate Petabytes of data daily.
• Processing: Sending all data to a central cloud would crash the network. Cities must use "Edge Computing" to process data locally at the source before sending summaries.
• Privacy: Handling constant surveillance and private location data securely is a major citizen concern.

2. Cost-Efficiency Challenges

• Massive Initial Investment: Digging up roads to install sensors and building control rooms costs billions of dollars.
• Maintenance Costs: Hardware breaks and gets damaged by weather or accidents, leading to ongoing financial burdens.
• Return on Investment (ROI): The benefits (like saving electricity) take many years to pay back the initial setup costs.

Q9. Trace the evolution of machine learning to modern deep learning architectures. Discuss the significance of regularization, batch normalization and dropout in training deep neural networks. (20 Marks)

1. Evolution: From ML to Deep Learning

• Early Machine Learning: Relied on algorithms like SVMs and Decision Trees. Required manual Feature Engineering (humans explicitly telling the computer what patterns to look for).
• The Problem: Failed to scale with massive datasets (Big Data) and complex tasks like image processing.
• Rise of Deep Learning: Powered by GPUs, deep networks (like CNNs and RNNs) perform Automatic Feature Extraction, figuring out important patterns from raw data without human help.

2. Significance of Key Training Techniques

• Regularization (L1 & L2 Penalty): Prevents "Overfitting" by adding a mathematical penalty to the network's weights, forcing the model to stay simple and learn general patterns rather than memorizing exact data noise.
• Batch Normalization: Prevents vanishing gradients by standardizing layer outputs before passing them to the next layer. This keeps numbers stable and allows for much faster training.
• Dropout: Randomly turns off a percentage of neurons during training. This prevents neurons from becoming "lazy" and forces all parts of the network to learn useful features independently, creating a highly robust model.

Q10. What are the core architectural components and service models of cloud computing? Discuss resource management, virtualization and security in cloud environments. How has the integration of Hadoop and big data technologies transformed cloud service delivery? (20 Marks)

1. Core Architectural Components and Service Models

• Core Architecture: Divided into the Front-End (user applications) and Back-End (massive data centers with servers and storage).
• Service Models:
  1. IaaS (Infrastructure): Renting raw virtual hardware.
  2. PaaS (Platform): Renting a pre-configured coding environment.
  3. SaaS (Software): Renting the final usable software application (e.g., Gmail).

2. Resource Management, Virtualization, and Security

• Virtualization: A "Hypervisor" slices a massive physical server into multiple independent "Virtual Machines" (VMs), allowing many customers to safely share one physical machine.
• Resource Management: Cloud is "elastic." Auto-scaling dynamically assigns more RAM/CPU during traffic spikes (like Black Friday) and scales down later.
• Security: Implemented via strict Data Encryption, Firewalls, and Identity Access Management (IAM) to protect data stored on third-party servers.

3. Hadoop, Big Data, and Cloud Transformation

• Before the Cloud: Analyzing Big Data meant spending millions buying and maintaining hundreds of physical servers.
• The Transformation: Cloud providers integrated Big Data tools directly into their platforms. Companies can now rent a massive Hadoop cluster for a few hours and shut it down.
• Impact: Shifted massive capital expenses to cheap operating expenses, democratizing AI and Big Data for small startups.