Boundary Adjustment and Mutation Analysis Tools

[ThomJ130]

This feature considers an previously known boundary which may have changed. For example, a machine learning model's region of validity has already been explored, but the model has since been further trained. How do we expand and contract the boundary to avoid unnecessarily re-exploring the search space? Furthermore, knowing how each boundary halfspace is adjusted will provide interesting insights into what effects the training had on the model.

Design Considerations

• Re-organize the halfspace data into two matrices: position matrix, $B$, and surface direction matrix, $O$.
• Represent adjustment weights as a column vector, $\vec d$
• New boundary position update function: $\hat B = B + O\cdot \vec d$
• Iteratively update the adjustment weights $\vec d$ until each vector in $B$ reaches a desired distance from the boundary.

[ThomJ130]

Furthermore, we will need to update the surface vector matrix after or while establishing the new boundary position matrix. This may be done using KNN to identify an orthogonal vector, but I am not sure yet.