Engineers use a three-step workflow to quickly train their AI agent.
- The engineer teaches the system what the AI should learn using a simple Domain Specific Programming Language called Inkling, which allows engineers to describe the characteristics of their specific simulation model and optimal behavior of the intelligent controller.
- The engineer connects the model with the service to establish the reinforcement learning loop by utilizing the Microsoft Simulink Toolbox. The Microsoft AI system provides input to the model, executes the simulation and reads back output for assessing the quality of the input with respect to the expected optimal control behavior. This is a one-time step that users will execute by using their local installation of Simulink.
- Microsoft’s Project Bonsai automatically scales simulation instances to reduce training times substantially. AI systems require large amounts of data samples and running simulation models at scale in parallel on the Azure cloud makes the system learn faster. Users just upload their model files and the Project Bonsai will do the rest.
The video provides a walk-through of the system and an overview of these steps. We will end up with a trained AI agent that has learned a policy that provides the correct action for any given state for keeping the pole upright.