MEPX 2023.02.16
12 views
Skip to first unread message
mihai....@gmail.com
Feb 18, 2023, 12:08:20 PM2/18/23
to Multi Expression Programming
Dear friends,
A new, major, version is MEPX has been released. Here is what is new:
1. Now, it is possible to solve symbolic regression with multiple outputs and multivariate time-series.
The number of problem's outputs can be set from Data |Training tab, by pressing the Num outputs button.
Currently, the fitness (the error) is calculated by summing the errors for each particular output. This can be a problem for cases where the outputs have different scales (e.g. output1 is in [0, 1] range and output2 is in [-10, 2000] range). I plan to improve this aspect by introducing Pareto dominance when comparing 2 solutions.
The folder with projects:
https://github.com/mepx/mepx-binaries
contains several new projects:
- symbolic regression/multiple outputs/building2.xml
this one is taken from PROBEN1, but with all outputs this time (consumption of electrical energy, hot water, and cold water),
- symbolic regression/multiple outputs/matrix_multiplication_10_0_1.xml
the formulas for 2x2 matrix multiplication are discovered here from data. The focus is only on finding formulas and not on finding the formulas with the minimum number of multiplications like in the Strassen algoritm.
- time series/multivariate/wage_growth.xml
Data are from Mehra, Y.P., Wage Growth and the Inflation Process: An Empirical Approach, 1994.
- time series/multivariate/Josephus_sequence.xml
finding the recursive formula for the Josephus sequence. Previously, with only one column of data, no formula was found. Now, a formula was found which depends on the index of the current number too.
When using the formula in Excel, make sure you select a consecutive number of cells equal to the number of inputs and then enter the formula and press Ctrl+Shift+Enter. For single output problems, only Enter must be pressed.
2. Also, now, the solutions (from a single run) are sorted by the number of utilized instructions too, not only by fitness. So, the displayed solution is always the one with minimum error and, among them, the one with minimum number of utilized instructions.
Best regards,
Mihai
A new, major, version is MEPX has been released. Here is what is new:
1. Now, it is possible to solve symbolic regression with multiple outputs and multivariate time-series.
The number of problem's outputs can be set from Data |Training tab, by pressing the Num outputs button.
Currently, the fitness (the error) is calculated by summing the errors for each particular output. This can be a problem for cases where the outputs have different scales (e.g. output1 is in [0, 1] range and output2 is in [-10, 2000] range). I plan to improve this aspect by introducing Pareto dominance when comparing 2 solutions.
The folder with projects:
https://github.com/mepx/mepx-binaries
contains several new projects:
- symbolic regression/multiple outputs/building2.xml
this one is taken from PROBEN1, but with all outputs this time (consumption of electrical energy, hot water, and cold water),
- symbolic regression/multiple outputs/matrix_multiplication_10_0_1.xml
the formulas for 2x2 matrix multiplication are discovered here from data. The focus is only on finding formulas and not on finding the formulas with the minimum number of multiplications like in the Strassen algoritm.
- time series/multivariate/wage_growth.xml
Data are from Mehra, Y.P., Wage Growth and the Inflation Process: An Empirical Approach, 1994.
- time series/multivariate/Josephus_sequence.xml
finding the recursive formula for the Josephus sequence. Previously, with only one column of data, no formula was found. Now, a formula was found which depends on the index of the current number too.
When using the formula in Excel, make sure you select a consecutive number of cells equal to the number of inputs and then enter the formula and press Ctrl+Shift+Enter. For single output problems, only Enter must be pressed.
2. Also, now, the solutions (from a single run) are sorted by the number of utilized instructions too, not only by fitness. So, the displayed solution is always the one with minimum error and, among them, the one with minimum number of utilized instructions.
Best regards,
Mihai
Reply all
Reply to author
Forward
0 new messages