Nowadays, most computers are based on a symbolic elaboration, that is, the problem is first encoded in a set of variables and then processed using an explicit algorithm that, for each possible input of the problem, offers an adequate output. However, there are problems in which resolution with an explicit algorithm is inefficient or even unnatural, for example with a speech recognizer; tackling this kind of problem with the classic approach is inefficient. This and other similar problems, such as autonomous navigation of a robot or voice assistance in performing an operation, are part of a very diverse set of problems that can be addressed directly through solutions based on reinforcement learning.
Reinforcement learning is a very exciting part of machine learning, used in applications ranging from autonomous cars to playing games. Reinforcement learning aims to create algorithms that can learn and adapt to environmental changes. To do this, we use external feedback signals (reward signals) generated by the environment according to the choices made by the algorithm. A correct choice will result in a reward, while an incorrect choice will lead to a penalization of the system. All of this is in order to achieve the best result obtainable.
The topics covered in this chapter are the following:
- An overview of machine learning
- Reinforcement learning
- Markov Decision Process (MDP)
- Temporal difference (TD) learning
- Deep Q-learning networks
At the end of the chapter, you will be fully introduced to the power of reinforcement learning and will learn the different approaches to this technique. Several reinforcement learning methods will be covered.