Amazon’s SageMaker has been updated in the drive to get more developers on the Machine-Learning train.
AWS has introduced SageMaker RL, which updates the cloud flinger’s Machine-Learning-as-a-service with toolkits for reinforcement learning.
The addition is built on top of the original managed Machine-Learning platform and provides containers for reinforcement learning with Apache MXNet and TensorFlow that include OpenAI’s Gym toolkit, Intel Coach and Berkeley Ray RLLib.
Reinforcement learning is an area of Machine Learning that lets a program or an agent learn by having it react to an environment by maximising rewards. Unlike other approaches, it does not start with a predefined, labeled dataset.
Instead, the environment is – most of the time, at least – simulated, the agent gets positive and negative feedback during interactions and – since the agent tries to maximise positive rewards – it starts to “learn” an optimal decision making strategy. Reinforcement learning is used in complex and somewhat unpredictable environments that can still be simulated – for example, in robotics and autonomous vehicles.
Simulations in SageMaker RL can be performed using commercial tools such as MATLAB and Simulink, first-party simulators like RoboMaker and Gym environments and environments developed using the Gym interface. Jupyter notebooks for the new offering are available on GitHub.
Third-party Machine Learning
Continuing the drive to attract developers, the AWS Marketplace has been updated to include a Machine Learning category featuring algorithms and model packages.
The digital catalogue for third-party software and services now contains a category that includes more than 150 machine learning algorithms and model packages. Sub-selections for different industries, such as manufacturing, media and retail are in place to help with a more focussed search.
The new category is meant to make Machine Learning more accessible by giving a sorted overview of options available. It should also keep customers from spending time developing approaches already in production someplace else.
To utilise one of the algorithms or packages you must pay a subscription and AWS resource fee. Offerings can be deployed from the SageMaker console or SDK, a Jupyter notebook or the AWS CLI.