Demystifying Supervised and Unsupervised Learning

Sylus

Welcome back. Today, we’re going to break down some of the fundamentals of machine learning, straight from the classroom.Let’s begin with supervised learning. This is one of the most common types of machine learning, where we train models using labeled data. Labeled means every input comes with the correct output attached. Think of it like a student learning under the supervision of a teacher. The student practices by seeing both the question and the right answer, and over time, learns how to predict the answer on their own.Now, the process of supervised learning usually goes like this: first, collect labeled training data. Next, split it into training and testing sets. Then decide which features matter most, pick an algorithm like a decision tree or a support vector machine, and train the model. Finally, you evaluate its accuracy using the test set.Supervised learning can be divided into two types. Regression is about predicting continuous values—like forecasting weather, predicting stock market trends, or estimating sales. Classification, on the other hand, deals with categories—yes or no, male or female, spam or not spam. Common algorithms here include logistic regression, random forests, decision trees, and support vector machines.Of course, supervised learning has both advantages and disadvantages. The big advantage is that, with enough examples, models can predict very accurately and solve real-world problems like fraud detection or spam filtering. The drawback is that it needs a lot of labeled data, and if the test data looks very different from the training data, performance drops.

Sylus

Now, let’s move to decision trees. A decision tree works like a flowchart. At the root, you start with all the data. Then, at each branch, you ask a question based on a feature—for example: is the weather sunny, overcast, or rainy? Depending on the answer, you move down the tree until you reach a leaf node, which gives you the decision or class.What decides the best place to split? That’s where concepts like entropy and information gain come in. Entropy measures disorder—how mixed the classes are. Information gain measures how much that disorder reduces when you split on a certain feature. The split that gives the highest information gain is chosen.But one problem with decision trees is overfitting—they can grow too complex and start memorizing noise in the data. To fix that, we use pruning—cutting away branches that don’t add real value. This makes the model simpler, faster, and better at generalizing to new data.Finally, let’s talk about unsupervised learning, where there are no labels. The system just looks for patterns in raw data. One of the most popular methods here is K-means clustering.Here’s how it works. First, you decide how many clusters, K, you want. Then you randomly select K points as cluster centers. Each data point is assigned to the nearest center. Next, you recalculate the cluster centers by taking the mean of all points in each cluster. You repeat this process—assigning and recalculating—until things stop changing.The goal is to make sure points inside the same cluster are very similar, and points in different clusters are as different as possible. This method is widely used in areas like data mining, image analysis, and bioinformatics.The main advantage of clustering over classification is that it adapts to changes in data and helps reveal useful hidden structures—even when you don’t know the categories in advance.

Sylus

Reinforcement learning is based on feedback. Imagine an agent—a computer program, or even a robot—that interacts with an environment. It takes an action, the environment responds, and the agent gets feedback in the form of a reward or penalty. Good actions earn positive rewards, bad actions bring negative ones. Over time, the agent learns the best strategy for achieving its goal.You can think of it like training a pet. A robotic dog, for instance, tries to move its legs. When it moves the right way, you reward it; when it stumbles, it gets no reward. Through trial and error, the dog eventually learns to walk smoothly.Now, let’s break down the key terms: -Agent: the learner or decision-maker. -Environment: the world the agent interacts with. -Action: the moves the agent can take. -State: the situation the agent is in after an action. -Reward: the feedback signal. -Policy: the agent’s strategy for choosing actions. -Value & Q-value: ways of estimating long-term benefits of actions, not just immediate rewards. There are different approaches too—value-based methods that try to maximize future rewards, policy-based methods that directly optimize decision strategies, and model-based methods where the agent builds a virtual model of the environment to plan its actions.

Sylus

Classification is a little different. Instead, it’s about picking which category something belongs to. Like sorting emails into “spam” or “not spam”—binary choices, or maybe even more than two. You could use a decision tree, a support vector machine, or something sleeker, but they’re all just mapping inputs into labeled boxes.

Sylus

Now, let’s make this practical with the Water Jug Problem, a classic puzzle.Here’s the setup: You have two jugs—one holds 3 liters, the other holds 5 liters. Your goal is to measure exactly 4 liters. The challenge is figuring out how to do it, step by step.We define the state space as a pair (x, y), where x is the water in the 3-liter jug and y is the water in the 5-liter jug. The initial state is (0, 0) because both jugs are empty. The goal is any state where one jug contains exactly 4 liters.The agent’s possible actions are:Fill a jug completely.Empty a jug.Pour water from one jug into the other, until either the first jug is empty or the second jug is full.By exploring different sequences—filling, pouring, emptying—the agent eventually learns the path that achieves the goal with the least wasted steps. Each correct step gives it a kind of ‘reward,’ each wrong step gives a penalty, and trial by trial, the agent discovers the solution.This puzzle is a simple way to see reinforcement learning in action: the agent explores, makes mistakes, but gradually learns the winning strategy.