# Types of Machine Learning Algorithms — Explained

Supervised, unsupervised, and others

Machine Learning informally defined as how the program improves itself gradually without manually/explicitly programmed. For example, an AI in email can identify which of the received messages is spam or not spam gradually.

The formal definition of Machine Learning is a computer program is said to learn (which is why it’s called “Machine Learning”) from experience E with respect to task T and some performance measure P, if its performance on T, as measured by P, improves with experience E. What does it mean? Lets return to the previous example, when an AI is observing us mark email as spam, it is said as the experience E, when the AI identify the spam emails, it is the task T. What is the performance measure? The performance measure (P) is emails that are correctly identified as spam. If the performance of T is increasing after E, that means we can infer that the AI is learning, hence, it is Machine Learning.

There are some types of Machine Learning algorithms, which I will explain on this article.

# Supervised Learning

At its core, supervised learning is basically how a machine can learn if we give it many inputs of dataset and observation with a certain label (that are already correct). The machine will attempt to learn the dataset and recognize new input and categorized it correctly.

One of the example of supervised learning is this housing price prediction.

Let’s say, given this data, we wanted to sell a house with a 750 feet². How much we can get from this 750 feet² house?

The learning algorithm — the supervised algorithm will attempt to make prediction by drawing a straight line over the graph. However, the straight line prediction isn’t forever, eventually, it will recognize another “good” prediction, like drawing a quadratic polynomial line or some other line that are closer to the shape of the graph.

The idea is, we give the “right answers” to the machine, and the machine will learn how the right answer is.

This housing price problem, is also called regression problem (which is one of the types of supervised learning), where we predict continous problem valued output.

Another example of supervised learning is the animal labelling problem.

Given a set of pictures of cats and dogs — and we already label the pics as either cat and dog, we give this dataset to the machine. After the data is processed, we test the machine by giving it a picture (still, it’s either a cat or dog), and let the machine decide it is dog or cat.

This type of problem is called classification problem (Discrete valued output), and it is one of the types of supervised learning.

Summary: The idea of supervised learning is making the machine learnt by giving it a dataset, the output we expect can be discrete (It usually about classification problem) or continue (It usually about regression problem)

# Unsupervised Learning

Unsupervised learning is when we doesn’t give the labels on the input we gives to the machine, we only give it the dataset and let it identify any structure of pattern inside. We don’t have any target or outcome to predict nor estimate. One of the uses of this type of algorithm is for segmenting customers based on a certain category.

Cocktail party problem

Let’s say we’re in a party where many people talk at the same time, obviously there are many overlapping voices and it’s hard to hear what the person in front of you.

Okay, maybe it was to extreme, let’s say there are only two person inside, with a two microphone, if we listen to the microphone, of course we willl hearing two distinct voices! This unsupervised learning will get rid of the other voice so we can hear the two distinct voices separately. This is one of the example of unsupervised learning.

# Reinforcement Learning

By using this type of algorithm, the machine will exposed to an environment where it trains itself with using a trial-and-error method. It will learn from pass experiences and capture the best knowledge to used.

The machine will be put in a game-like situation where the gets rewards or penalties (But not both) for the action it performs. The goal is to maximize the reward and penalties.

To explain it more simple, let’s take a look at this example.

The AI is the robot, the goal is to move to diamond with fewer steps as possible and avoiding fire. Each step will give reward for the robot (If the step is right) and wrong step will reduce it. When it reach the diamond, the total reward will be calculated.

Important elements in reinforced learning:

• Input: Initial state of the model
• Output: Variety of the solutions, effect of the input to the machine
• Training: Based on the input, the model will return a state and the AI will be rewarded/punished based on the output
• Best solution: Solution that make up for the most reward
• One last thing (And its quite obvious): The model will learn gradually

AI/ML

Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot