Unit 5: Dimensionality Reduction, Genetic Algorithms & Reinforcement Learning

 

Machine Learning Techniques (MCA556)

From your syllabus.

---

Dimensionality Reduction

In Machine Learning, datasets may contain many features (columns).

Example:

Student Data
------------
Name
Age
Gender
Address
Attendance
Marks
Projects
Activities
...

Too many features can:

• Increase training time
• Increase memory usage
• Cause overfitting

Dimensionality Reduction reduces the number of features while preserving important information.

---

Benefits

• Faster computation
• Less storage
• Better visualization
• Reduced overfitting

---

Principal Component Analysis (PCA)

Most important dimensionality reduction technique.

Purpose:

• Convert many features into fewer important features.

Idea:

• Preserve maximum variance.
• Reduce dimensions.

Example:

100 Features
     ↓
PCA
     ↓
10 Important Features

Applications:

• Face Recognition
• Image Compression
• Data Visualization

---

Linear Discriminant Analysis (LDA)

Used for:

• Dimensionality Reduction
• Classification

Difference:

PCA	LDA
Unsupervised	Supervised
Maximizes variance	Maximizes class separation

Applications:

• Face recognition
• Pattern classification

---

Factor Analysis

Statistical method used to identify hidden factors affecting data.

Example:

Student Performance depends on:

• Intelligence
• Study Hours
• Motivation

These hidden variables are called factors.

Applications:

• Psychology
• Market Research
• Social Sciences

---

Independent Component Analysis (ICA)

Separates mixed signals into independent components.

Example:

Two people speaking simultaneously.

ICA can separate:

Mixed Audio
      ↓
ICA
      ↓
Speaker 1
Speaker 2

Applications:

• Signal Processing
• Medical Data Analysis
• Audio Separation

---

Locally Linear Embedding (LLE)

Non-linear dimensionality reduction technique.

Assumption: Nearby data points remain nearby after transformation.

Used when data lies on a curved surface.

Applications:

• Pattern recognition
• Data visualization

---

Isomap

Isomap = Isometric Mapping

Advanced dimensionality reduction technique.

Purpose:

• Preserve geometric structure of data.

Applications:

• Image analysis
• Visualization
• Pattern recognition

---

Least Squares Optimization

Used to minimize prediction error.

Idea:

Find the best line that minimizes squared errors.

Linear Regression is based on Least Squares Optimization.

---

Evolutionary Learning

Inspired by biological evolution.

Key concepts:

• Selection
• Mutation
• Crossover
• Survival of the fittest

Used to solve optimization problems.

---

Genetic Algorithms (GA)

One of the most important evolutionary algorithms.

Inspired by natural selection.

---

Basic Terminology

Chromosome

A possible solution.

---

Population

Collection of chromosomes.

---

Fitness Function

Measures solution quality.

Higher fitness = Better solution.

---

Genetic Algorithm Steps

Step 1

Initialize Population

Generate random solutions.

---

Step 2

Evaluate Fitness

Check quality of each solution.

---

Step 3

Selection

Choose best solutions.

---

Step 4

Crossover

Combine parents to create offspring.

Parent A + Parent B
          ↓
       Child

---

Step 5

Mutation

Randomly modify genes.

Purpose:

• Maintain diversity

---

Step 6

Replacement

Create next generation.

Repeat until optimal solution found.

---

Applications of Genetic Algorithms

• Scheduling
• Route Optimization
• Machine Learning
• Robotics
• Engineering Design

---

Reinforcement Learning (RL)

Learning through rewards and punishments.

Agent learns by interacting with environment.

---

Components of RL

Agent

Learner.

Example: Robot

---

Environment

World around the agent.

Example: Road

---

Action

Decision taken by agent.

Example: Move Left

---

Reward

Feedback received.

Example:

Correct Action → +10
Wrong Action → -5

---

Reinforcement Learning Process

Agent
  ↓
Action
  ↓
Environment
  ↓
Reward
  ↓
Learning

---

Applications of Reinforcement Learning

• Self-driving cars
• Robotics
• Game playing AI
• Resource management

---

Markov Decision Process (MDP)

Mathematical framework for Reinforcement Learning.

An MDP contains:

1. State (S)
2. Action (A)
3. Reward (R)
4. Transition Probability (P)

---

Example of MDP

Robot Navigation:

State:
Current Position

Action:
Move Left / Right

Reward:
Reach Destination

Next State:
New Position

---

Markov Property

Future state depends only on the current state.

Not on previous history.

---

Important Exam Questions

Short Questions

1. What is PCA?
2. Define LDA.
3. What is ICA?
4. Define Isomap.
5. What is Genetic Algorithm?
6. What is Fitness Function?
7. Define Reinforcement Learning.
8. What is MDP?

---

Long Questions

1. Explain PCA with advantages.
2. Differentiate PCA and LDA.
3. Explain Genetic Algorithm with steps.
4. Discuss Evolutionary Learning.
5. Explain Reinforcement Learning architecture.
6. Explain Markov Decision Process.

---

Quick Revision

• PCA = Reduce dimensions while preserving variance.
• LDA = Reduce dimensions while separating classes.
• ICA = Separate mixed signals.
• Isomap = Preserve geometric structure.
• GA = Optimization inspired by evolution.
• Population = Collection of solutions.
• Fitness Function = Quality measure.
• RL = Learning through rewards.
• MDP = Mathematical model for RL.
• Markov Property = Future depends only on current state.

✅ Machine Learning Techniques (MCA556) is now complete.

Next Subject Options

1. .NET Framework with C# (MCA552)
2. Compiler Design (MCA554)
3. Optimization Techniques (MCA555)
4. Advanced JavaScript (MCA557 B/C)

For exams, I would suggest Compiler Design next because it is usually considered the toughest paper and benefits from early preparation.