search

ethz-adrl / ifopt

An Eigen-based, light-weight C++ Interface to Nonlinear Programming Solvers (Ipopt, Snopt)
http://wiki.ros.org/ifopt
BSD 3-Clause "New" or "Revised" License
774 stars 156 forks source link

Question about warm starting #67

Closed junhyeokahn closed 3 years ago

junhyeokahn commented 3 years ago

Hello! I appreciate your efforts in maintaining this nice repo.

I have a quick question about the warm starting of the optimizer. Can I just set variables with the numbers I want to initialize as you do in here? Or should I call extra method such as SetOption("warm_start_init_point", "yes");?

awinkler commented 3 years ago

Hi! It works without the option, although IPOPT might do some smarter tricks when you additionally set that option, I haven't experimented with that yet.

Generally I've found it's quite difficult to warm-start interior point methods like IPOPT, because of the way they solve the problem: The initial variables are often changed a lot to move into the center of the constraints during the first iteration. SQP methods, like SNOPT, might be a better choice if you have trustworthy information of the initial values (like in Model-Predictive-Control, where the initial values can be the solution of the previous solve).

junhyeokahn commented 3 years ago

Thanks for your kind reply!

mzwang828 commented 3 years ago

Hi! It works without the option, although IPOPT might do some smarter tricks when you additionally set that option, I haven't experimented with that yet.

Generally I've found it's quite difficult to warm-start interior point methods like IPOPT, because of the way they solve the problem: The initial variables are often changed a lot to move into the center of the constraints during the first iteration. SQP methods, like SNOPT, might be a better choice if you have trustworthy information of the initial values (like in Model-Predictive-Control, where the initial values can be the solution of the previous solve).

Hi @awinkler, I have the same question about using SNOPT. Do I just set the variables with good initial values or I should also change the parameter "Cold" in the snopt_sovler.cc? Once I set Cold = 2, which stands for warm starting, SNOPT gives warnings of "Invalid argument Fstate" and it seems that Fstate are not assigned with any meaningful values in SnoptAdapter. Thank you!