Solving problems with machine learning

Among the different machine learning approaches, there are three main ways of learning, as shown in the following list:

• Supervised learning
• Unsupervised learning
• Reinforcement learning

Given a set of example inputs X, and their outcomes Y, supervised learning aims to learn a general mapping function f, which transforms inputs into outputs, as f: (X,Y).

An example of supervised learning is credit card fraud detection, where the learning algorithm is presented with credit card transactions (matrix X) marked as normal or suspicious (vector Y). The learning algorithm produces a decision model that marks unseen transactions as normal or suspicious (this is the f function).

In contrast, unsupervised learning algorithms do not assume given outcome labels, as they focus on learning the structure of the data, such as grouping similar inputs into clusters. Unsupervised learning can, therefore, discover hidden patterns in the data. An example of unsupervised learning is an item-based recommendation system, where the learning algorithm discovers similar items bought together; for example, people who bought book A also bought book B.

Reinforcement learning addresses the learning process from a completely different angle. It assumes that an agent, which can be a robot, bot, or computer program, interacts with a dynamic environment to achieve a specific goal. The environment is described with a set of states and the agent can take different actions to move from one state to another. Some states are marked as goal states, and if the agent achieves this state, it receives a large reward. In other states, the reward is smaller, non-existent, or even negative. The goal of reinforcement learning is to find an optimal policy or a mapping function that specifies the action to take in each of the states, without a teacher explicitly telling whether this leads to the goal state or not. An example of reinforcement learning would be a program for driving a vehicle, where the states correspond to the driving conditions, for example, current speed, road segment information, surrounding traffic, speed limits, and obstacles on the road; and the actions could be driving maneuvers, such as turn left or right, stop, accelerate, and continue. The learning algorithm produces a policy that specifies the action that is to be taken in specific configurations of driving conditions.

In this book, we will focus on supervised and unsupervised learning only, as they share many concepts. If reinforcement learning sparked your interest, a good book to start with is Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew Barto, MIT Press (2018).