ASSEMBLE is a framework grant from SSF - the Swedish Foundation for Strategic Research. The project will last for five years starting 2016. The level of funding is 5.8 MSEK per year. ASSEMBLE stands for Automating system specific model-based learning.
The aim of the ASSEMBLE project is to automate probabilistic modeling of dynamical systems (and their surroundings) via a formally defined probabilistic modeling language. The proposed modeling language will offer a valuable abstraction for the users, allowing them to focus on their specific problems and not having to spend time on the underlying learning algorithms. For users this will save huge amounts of time, since they can instead maintain focus on the actual problem they are trying to solve. At the same time our language will provide a way for algorithm developers to make their new developments reach a big group of users much faster than today. In this sense ASSEMBLE will provide a market place for algorithm developers and their users. The following figure illustrates our approach.
The application model is what will provide the abstraction that allows system designers to express their problems and constraints. We will herein focus on probabilistic models as they provide a systematic way to express and work with the uncertainty inherent in most settings. There will be one or more learning algorithms that can be used to learn model specific variables from data for any given model. The application model will be described in a modeling language and automatically combined with suitable inference methods though a probabilistic model compiler and thereby yield computationally efficient application specific machine learning solutions.
Together with our partner companies and endusers (Autoliv, Greenely, Karolinska, and ABB), we have selected demonstrators that drive our research and safeguard the project’s direct relevance for the Swedish industry. These demonstrators involve understanding the environment surrounding a car, energy disaggregation using smart meters, automated analysis of cell migration, energyaware computing, and container crane automation.
While these demonstrators cover a diverse range of industrial expertise, each of them involves a dynamic system that interacts with an environment perceived through a range of sensors. At the core, we have models and inference methods that turn sensor data into information about the system’s behavior and its environment. This information is passed on to humans or to the system itself, which takes appropriate actions. By using the new modeling language, we will automate this process.