When you do something well you've probably been rewarded for it. Reinforcement learning is a machine learning technique which uses the same reward-based approach. Reinforcement learning involves an agent which takes particular actions within an environment in such a way as to always maximize a goal. It's always difficult to first comprehend something that has been explained in the abstract or generalized form, so let's provide a real world example. 

You're teaching a bot of sorts to play the arcade game PAC-MAN. The goal in PAC-MAN is to get as many points as possible. Points are accumulated by eating dots and catching ghosts when powered up. Using reinforcement learning, the agent is the bot and the environment is the game layout in which PAC-MAN is moved about in. Each time the bot moves PAC-MAN for which a dot is eaten, a reward is provided. During any time that PAC-MAN hasn't powered up, PAC-MAN must avoid being caught by any ghost. Being caught is a punishment of sorts. Using reinforcement learning, the bot will learn how to play the complete game without being explicitly programmed to do so. And this is really the critical take-away. In essence, reinforcement learning has been used to learn a strategy to play the game. 

To help with reinforcement learning, Amazon SageMaker RL has been introduced as a part of SageMaker. Using Amazon SageMaker RL you can build and create advanced reinforcement learning-based models without having to manage or provide any of the underlying infrastructure. Interestingly, Amazon SageMaker RL can be used to create RL models that are then used with AWS DeepRacer and/or AWS RoboMaker services, as well as in many other IoT type scenarios. 

Amazon SageMaker RL is preloaded with several reinforcement learning toolkits and frameworks. Of notable mention are Intel Coach and Ray RL, both specifically created for reinforcement learning, TensorFlow and MXNet which are used to support deep learning-based RL, and Open AI Gym which provides simulation environments in which RL can take place. 

The general approach for creating a reinforcement learning model using SageMaker involves setting up a Markov Decision Process, an MDP and an MDP consists of the following sequence of steps: an objective, an environment, state, action and reward. When performing SageMaker RL training, MDP sequence is repeated many times referred to as episodes. Next you define the environment in which the reinforcement learning will take place. Often this is a simulation and prepackaged environment and can be imported from Open AI Gym. Training code in the form of a Python script is uploaded and now you're ready to kick off the training and wait for the results. Often you will plot training metrics using notebooks supported in SageMaker and this is done using the Python SDK analytics library. Visualizing metrics is important as will it render how the reward accumulation is improving over time. Additionally, custom metrics that you configure within CloudWatch can be utilized to gain insights into the RL training process. Finally once the reward accumulation has surpassed a level of your choosing, you are ready to deploy the reinforcement learning model.