Take a Ver 11 program that uses Levenburg Marquardt, and initializes all optimization parameters with real numbers. Convert the program to V14. Run it in 14. It answers with complex results in the optimized parameters. This is a bug, I think...or is there a new way to control the optimization results?
