You may be familiar with the notion of supervised learning, which is the most studied and well-known machine learning problem. Its basic question is: how do you automatically build a function that maps some input into some output, when given a set of example pairs? It sounds simple in those terms, but the problem includes many tricky questions that computers have only recently started to deal with some success. There are lots of examples of supervised learning problems, including the following:

These questions can look different, but they share the same idea: we have many examples of the input and desired output, and we want to learn how to generate the output for some future, currently unseen inputs. The name, supervised comes from the fact that we learn from the known answers, which were obtained from some supervisor who has provided us with those labeled examples.

At the other extreme, we have the so-called unsupervised learning, which assumes no supervision that has no known labels assigned to our data. The main objective is to learn some hidden structure of the dataset at hand. One common example of such an approach to learning is the clustering of data. This happens when our algorithm tries to combine data items into a set of clusters, which can reveal relationships in data.

Another unsupervised learning method that is becoming more and more popular is, Generative Adversarial Networks (GANs). When we have two competing neural networks, the first of them is trying to generate fake data to fool the second network, while the other is trying to discriminate artificially generated data from data sampled from our dataset. Over time, both of them are becoming more and more skillful in their tasks by capturing subtle specific patterns of your dataset.

RL is the third camp and lays somewhere in between full supervision and a complete lack of predefined labels. On the one hand, it uses many well-established methods of supervised learning such as deep neural networks for function approximation, stochastic gradient descent, and backpropagation, to learn data representation. On the other hand, it usually applies them in a different way.

In the next two sections of the chapter, we'll have the chance to explore specific details of the RL approach including its assumptions and abstractions in its strict mathematical form. For now, to compare RL to supervised and unsupervised learning, we'll take a less formal, but more intuitive description.

Imagine you have an agent that needs to take actions in some environment. A robot mouse in a maze is a good example, but we can also imagine an automatic helicopter trying to make a roll, or a chess program learning how to beat a grandmaster. Let's go with the robot mouse for simplicity.

Figure 1: Robot mouse maze world

Its environment is a maze with food at some points and electricity at others. The robot mouse can take actions such as turn left/right and move forward. Finally, at every moment it can observe the full state of the maze to make a decision about the actions it may take. It is trying to find as much food as possible, while avoiding an electric shock whenever possible. These food and electricity signals stand as a reward given to the agent by the environment as additional feedback about the agent's actions. The reward is a very important concept in RL, and we'll talk about it later in the chapter. For now, it will be enough to understand that the final goal of the agent is to get as much total reward as possible. In our particular example, the mouse could suffer a bit of an electric shock to get to the place with plenty of food—this will be a better result for the mouse than just standing still and gaining nothing.

We don't want to hard-code knowledge about the environment and the best actions to take in every specific situation into the robot—it will take too much effort and may become useless even with a slight maze change. What we want to do is to have some magic set of methods that will allow our robot to learn on its own how to avoid electricity and gather as much food as possible.

Reinforcement Learning is exactly this magic toolbox, which plays differently from supervised and unsupervised learning methods. It doesn't work with predefined labels as supervised learning does. Nobody labels all the images the robot sees as good or bad or gives it the best direction to turn in.

However, we're not completely blind as in an unsupervised learning setup—we have a reward system. Rewards can be positive from gathering the food, negative from electric shocks, or neutral when nothing special happens. By observing such a reward and relating it to the actions we've taken, our agent learns how to perform an action better, gather more food, and get fewer electric shocks.

Of course, RL generality and flexibility comes with a price. RL is considered to be a much more challenging area than supervised and unsupervised learning. Let's quickly discuss what makes Reinforcement Learning tricky.

The first thing to note is that observation in RL depends on an agent's behavior and to some extent, it is the result of their behavior. If your agent decides to do inefficient things, then the observations will tell you nothing about what they have done wrong and what should be done to improve the outcome (the agent will just get negative feedback all the time). If the agent is stubborn and keeps making mistakes, then the observations can make the false impression that there is no way to get a larger reward—life is suffering—which could be totally wrong. In machine learning terms, it can be rephrased as having non-i.i.d data. The abbreviation i.i.d stands for independent and identically distributed, a requirement for most supervised learning methods.

The second thing that complicates our agent's life is that they need to not only exploit the policy they have learned, but to actively explore the environment, because, who knows, maybe by doing things differently we can significantly improve the outcome we get. The problem is that too much exploration may also seriously decrease the reward (not to mention that the agent can actually forget what they have learned before), so, we need to find a balance between these two activities somehow. This exploration/exploitation dilemma is one of the open fundamental questions in RL.

People face this choice all the time: should I go to an already known place for dinner or try this new fancy restaurant? How frequently should you change jobs? Should you study a new field or keep working in your area? There are no universal answers to these questions.

The third complication factor lays in the fact that reward can be seriously delayed from actions. In cases of chess, it can be one single strong move in the middle of the game that has shifted the balance. During learning, we need to discover such casualties, which can be tricky to do over the flow of time and our actions.

However, despite all these obstacles and complications, RL has made huge improvements over recent years and is becoming more and more active as a field of research and practical application.

Interested? Let's get to the details and look at RL formalisms and play rules.