FormaK aims to combine symbolic modeling for fast, efficient system modelling with code generation to create performant code that is easy to use.
This design focuses on "Integration to scikit-learn to leverage the model selection and parameter tuning functions". More specifically, this design focuses on using scikit-learn tooling to automatically select the innovation filtering level from data.
The promise of this design is that all parameters could be selected automatically based on data instead of requiring hand tuning; however, this design will focus narrowly on selecting the innovation filtering level as a motivating example.
Pull Request: #21 Commit: 5ce60af
Why?¶
The Kalman Filter has a number of selectable parameters: - process noise for each reading of the state vector - sensor noise for each element of the reading vector
The innovation filtering approach also introduces an additional parameter and future designs will likely introduce additional parameters as well.
This could be treated as a magic number or a human-tuned parameter; however, machine learning provides a better process: cross validation.
Given data and a measure of fit (innovation), we can select different sets of parameters based on training data, validate it against other training data and cross-validate it against a held-out test set. This means that data can point us to the correct parameters for the filter and the cross validation approach should reduce the risk of overfitting.
All of this is of some complexity to set up and get running, so integrating it into FormaK will allow for easily expanding its use and re-use in my own work, sharing the process with other people without having to re-implement the process.
What's next?¶
As a next step, I'm opening up the roadmap for FormaK via Github Issues. For each proposal, I'm outlining the overview of the concept, how it fits into the FormaK project goals, design decisions that will need to be made and seeking feedback.
After getting that part of the project set up, I'll move onto the next feature. I suspect the next feature will be implementing some portion of the netcode enhancement, but that's still to-be-determined.