exploit duality in Symbolic Regression

[remiadon]

One idea stolen from Linear Optimisation is duality
Applying this to Symbolic Regression, one can image a setup where at generation G:

• best individuals in are picked, for further offsping (traditional definition of genetic algos)
• worst indiduals are picked, and most common subepxressions are extracted. While those badly performing expressions are used to form a next generation (maximizing the loss), the most common subepxressions are blacklisted from the mating process of the best performing ones.