Reinforcement learning

The need of an alternative to traditional learning techniques arose with the design of the first autonomous systems.

The problem

Autonomous systems are semi-independent systems that perform tasks with a high degree of autonomy. Autonomous systems touch every facet of our life, from robots and self-driving cars to drones. Autonomous devices react to the environment in which they operate. The reaction or action requires the knowledge of not only the current state of the environment but also the previous state(s).

Autonomous systems have specific characteristics that challenge traditional methodologies of machine learning, as listed here:

• Autonomous systems have poorly defined domain knowledge because of the sheer number of possible combinations of states.

• Traditional nonsequential supervised learning is not a practical option because of the following:

  • Training consumes significant computational resources, which are not always available on small autonomous devices

  • Some learning...