Closed timm closed 7 years ago
Maybe a more constructive approach: For a particular specified solution, how can we assess the relative goodness? Is this solution sitting alone in a trough, and a small change might make the solution much better? Or is it sitting on a plateau with no objective difference from its neighbors, and so small changes don't have much impact?
Do we have a measure of the "distance" between two solutions, e.g., how many decisions are different?
so some distance measure that is not just the syntactic delta between items, but also some measure of the effect on the goals of jumping between them
george.. is the b^2/(b+r) distance insightful here?
The bayesian support(b^2/b+r) looks like a good criterion to categorize the decisions.
There are two measures here: one is the relative goodness of a solution, the other is a measure of how different is the path that gets us to two adjacent solutions. If changing one constituent decision turns a bad solution into a significantly better solution, that could be a good thing in the process.
@bigfatnoob : dont implement this one (unless SEI wants it). its too evil.
not matter what, X has to be in the final solution. move heaven and hell to make it so.