How do the duals get initialized?

[mahonbt1]

We are seeing minor differences in iterations between different builds of our solver that should be exactly the same. The best we can tell is that the Dual variables are going uninitialized, or something to that effect.

I wanted to ask how init_lambda and lambda work inside of the TNLP's get_starting_point() method. Is there an option we should be setting at startup to get this under control?

[svigerske]

By default, Ipopt will calculate some initial value for the dual values. It uses a least-square estimate for constraint multipliers and a constant for bound multipliers, see the various Initialization options.

If you enable warm_start_init_point, then Ipopt will call get_starting_point() with init_lambda = true and you are supposed to fill lambda with a suitable value.