I want to minimize a system of equation that has multiple right hand side (RHS). In the following piece of code, I defined my minimization function, which takes different RHS to compute the minimized solution and I have to loop through the entire RHSs.
For the problem I am dealing with ntime = 1000, n = 4000, and m = 50. x is different for each timeframe corresponding to its RHS. In other words, the solution x_all is of size ntime×n. Since the only thing that changes in the above code is the RHS, is there anyway to input all the RHSs to CVX and eliminate the for-loop and make the whole process more efficient?
I don’t know whether the following, which involves only a single pass through cvx_begin … cvx_end., would be more or less efficient than your current code.
Presuming that you have enough memory, stack all ntime x vectors into one x, stack the b vectors, form matrices with A and L on the block diagonals, and use these in a single objective function instead of your current formulation.
Alternatively, you could build up the objective function as an expression in a for loop, which adds the objective functions for each RHS, as suggested by me at Large measurements b matrix .
Either way, I believe you will be solving what decomposes into ntime “independent” optimization problems, which are the same as your current problems. But it will be submitted as a single problem.
Thanks Mark. I had tried augmenting all the RHS together to solve a huge system of equations, which was not faster. Also seems like summing objecting functions (as suggested by you in the link) is not faster either. Solving for each timestep for my case looks like the fastest among these methods.
If I use CVX in other environments rather than MATLAB, would it help?
Sounds very interesting to see what CVX passes to the solver. However, I was not able to find any information on solver_dump option in the documentation. I also found this thread . Would you post an example how solver_dump works?