Reinforcement learning is when you show the algorithm by presenting examples without labels like in unsupervised learning. It is possible to accompany an example with either positive or negative feedback, based on the solution proposed by the algorithm. Reinforcement learning is linked to applications in which the algorithm is required to make decisions (so it is not only descriptive, like in unsupervised learning) and the choices have consequences. In the world of humans it's just like the process of learning through trial and failure.
Learning is enhanced when mistakes are made because they carry a cost (cost in time, regret, loss of time hurt etc.) which teaches you that a specific course of decision is more likely to work than other. A great instance of reinforcement learning is when computers are taught to play video games without any assistance.
In this scenario an application demonstrates the algorithm by presenting instances of specific scenarios like having the gamer trapped in a maze and trying to avoid an opponent. The program allows the algorithm to know the results of the actions it makes and it learns while trying to stay clear of what it finds to be risky and also to ensure that it is able to survive.
Take a glimpse of how Google DeepMind has created a reinforcement learning program that can play old videogames from Atari. While watching the video, you will notice that the program starts out inexperienced and unprofessional, but it gradually gets better through Machine Learning Training until it is an elite player.
Top comments (0)