]Things to know about the offline development of additive modules
Steps/recommendations:
- Primary OAM (in case of large quantity of data, derived as a pivot-table, where objects are in the rows, their attributes and measurement units are in the columns, and the far right column is the Y.)
- Ranking OAM (measurement unit: rank) is formed manually, with the help of the RANKING() and PERCENTRANK () functions.
- Definition of an empty stair-function (measurement unit is equal with the one in the Y column, in the row header, there are numbers between 1 and n, where n is the number of stairs, and it is equal with the highest rank).
- In certain cases, stairs shall be initialized (e.g. with values descending from n to 1) in order to make the solver run.
- Definition of the COCO-matrix (measurement unit is the sameas the Y's, objects are in the row header, the VLOOKUP function looks for the stair values in the stair table, which is calculated by the solver, based on the rankings table).
- The COCO matrix is the objective function of the solver, so the estimation values are needed to be created here, which, in case of additive models, is the sum of the input stair levels per objects.
- The functions that determine the differences of estimations and facts per objects, and which calculates the square sum of this vector, shall be added here.
- If we want to add parameters to the solver quickly, then an auxiliary matrix shall be made too (same measurement unit as the Y's), which shall be the difference of two consequtive elements of the columns of the stair matrix.
- Functions helping interpretation: the sum of the first stair level (genetic/Y, measurem is same as the Y's) can be interpreted as potential.
- The proportion of the sum of columns in the COCO-matrix and the sum of the estimated values means the weight (%) of the attributes.
- The sum of the estimation and facts should be equal. If it is not, then the proportion of the difference must be projected for each stair steps.
- Solver parameters: target cell is the cell which contains the (square) sum. Changing cells are the (empty) stair steps. Constraints: the value of the stairs is higher than/equal with (if necessary, and in each column the better stair is a better place) cannot worth less than the worse OR the auxiliary matrix's each cell is higher than/equal with zero.
- The solver needs multiple subsequent runs, always accepting the results of the previous run.
