tag:blogger.com,1999:blog-593563533834706486.post7234275334003709543..comments2024-05-21T03:18:48.423-04:00Comments on Yet Another Math Programming Consultant: Choosing between NLP solvers: interior point or active set.Erwin Kalvelagenhttp://www.blogger.com/profile/09496091402502236997noreply@blogger.comBlogger4125tag:blogger.com,1999:blog-593563533834706486.post-25972344571148285812023-03-16T17:12:22.067-04:002023-03-16T17:12:22.067-04:00Consider for example the Lagrangian relaxation of ...Consider for example the Lagrangian relaxation of a combinatorial problem. One optimizes a black-box function which consists in solving an optimisation subproblem. Yet, the sub-gradient is directly available without any additional costAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-593563533834706486.post-31872219690737024352023-03-16T14:01:36.079-04:002023-03-16T14:01:36.079-04:00Black-box optimization is really synonym to DFO.Black-box optimization is really synonym to DFO.Erwin Kalvelagenhttps://www.blogger.com/profile/09496091402502236997noreply@blogger.comtag:blogger.com,1999:blog-593563533834706486.post-32846204520189594982023-03-16T08:50:40.116-04:002023-03-16T08:50:40.116-04:00And out of curiosity, could you share the exact mo...And out of curiosity, could you share the exact model you used for the experiments?<br /><br />"Expensive function evaluations are the rule when calling a function that will run some simulation. In this case, it surely is a bad idea to use finite differences. That would create a large number of extra function evaluations. Rather we want to use some DFO (Derivative-Free Optimization) method [3]."<br />It happens that an expensive black-box function still provides the derivatives without additional cost. In this case, one can expect derivative-based algorithms to work better than DFO algorithmsAnonymousnoreply@blogger.comtag:blogger.com,1999:blog-593563533834706486.post-1899848595907980172023-03-16T08:39:02.772-04:002023-03-16T08:39:02.772-04:00You could modify the problem a bit in the second e...You could modify the problem a bit in the second experiment before resolving to make the comparison more relevant. For example, changing a bound of a variable with a non-integral value to simulate what happens in a branch-and-bound algorithm.Anonymousnoreply@blogger.com