Source: Deep Learning on Medium
Trial and Error ordinarily is the means by which humans and other organisms learn and progress over time. In the pursuit of anthropoid machines, 2019 is the year where Deep Reinforcement Learning lifts off.
For any problem where you can define an environment, states in the environment, actions and rewards for every state-action pair, that problem can be solved with Reinforcement Learning.
For example, if you want your rover to navigate from point A to point B on the surface of Mars, you start off by defining an environment. For simplicity, consider a grid of 5×5.
There are several ways for the rover to find water, but if you discount the reward you receive for each additional step taken, give negative rewards when the rover comes in touch with a volcano or an impact crater, you can create an algorithm to find water as quickly as possible!
And this is reinforcement learning. Ciao.