Extending Microservice Model Validity using Universal Differential Equations
Albin Heimerson, Johan Ruuskanen · IFAC-PapersOnLine · 2023
When creating models of a system, there is always a tradeoff between the ease of modelling a part and the increased value it brings to the model. Learning a model using machine learning, we might have less control over what dynamics to capture, but can also capture things we don't necessarily understand. Using universal differential equations we can combine the two, taking scientific models and embedding machine learning into them, with the goal of giving us the best of both worlds. In this paper, we extend an existing model with small neural networks to capture missing dynamics. The specific use-case involves a microservice fluid model, where the learned extension improves the range of parameters the combined model can reliably produce predictions over. This means the model can be reused in a wider parameter space before needing to be retrained. We also explore the possibility of imposing bias on the network based on features of the model, and how that affects performance.