Machine Learning is an AI technology that learns and continuously improves performance without explicit programming. Machine learning algorithms establish a mathematical model for a specific field and complete this model by training with data, enabling them to predict outcomes or make decisions.
Concept of Machine Learning
Machine learning techniques include unsupervised learning, supervised learning, and reinforcement learning.
Unsupervised learning groups data with similar characteristics based on data features. There is no reference data regarding the learning objective in unsupervised learning. For example, when classifying customer segments based on purchasing data, unsupervised learning can find groups of customers with similar purchasing patterns. Since there is no data indicating which customers are valuable and which are not, a business expert would need to analyze the categorized groups to distinguish valuable from non-valuable customers. In other words, unsupervised learning performs the initial classification, and a business expert conducts a secondary evaluation.
A representative algorithm used in unsupervised learning is the k-means algorithm. Here, k represents the number of groups to classify, so if k=5, the algorithm classifies customers into five groups. The group center (k) is set arbitrarily, and members are regrouped continuously until the average distance between each member and the center is minimized.
Supervised learning is a technique where data with clear outcomes (targets) is used to complete a mathematical model, which then predicts results when new data is introduced. For example, suppose we want to create a model that predicts stock prices. The data required for training would include factors that affect stock prices on each date and the stock prices themselves. The goal of the training is to predict the price for an unknown date given influencing factors, hence the term "supervised learning." After completing the model, we can input basic data to predict the stock price for the following day.
The classic example of supervised learning is the problem of classifying images as dogs or cats. If we want to create a model that can identify whether an image is a dog or a cat, we must train it on millions of labeled images. Once trained, the model will classify new images as either dogs or cats on its own.
Reinforcement learning differs somewhat from supervised and unsupervised learning in concept. Reinforcement learning uses rewards based on the agent’s actions and state. Although this was explained earlier, let's review the concept briefly.
Consider a baby learning to walk. The baby doesn’t learn from anyone else. If the baby shifts weight onto the left leg and tries to stand up on the right side but falls painfully, they may try to stand up on the left side next time. When standing successfully, the baby receives praise and a broader view, which is a reward for standing correctly. Although the baby can’t get up easily the first time, they may stand a bit more easily, motivated by the sense of achievement.
The same goes for walking. If lifting the left foot while shifting weight off the right leg results in pain from falling, the baby will try supporting the right leg when lifting the left. The goal is to avoid pain and reach the destination quickly. Through falling, hurting, and being praised, the baby learns to walk.
Reinforcement learning models this process mathematically. The most fundamental concept is MDP (Markov Decision Process). Reinforcement learning theory is based on MDP, so understanding MDP is essential to understanding reinforcement learning. We will cover reinforcement learning and MDP in more detail shortly.
Since the goal of this book is to understand reinforcement learning, we will focus on supervised learning, which is more closely related to reinforcement learning, and omit an extensive discussion of unsupervised learning.
