[SCIP] Large B&B trees with Vanillafullstrong and idempotent=True

[Caroline Spieckermann]

Hello,


I'm trying to figure out which (vanilla) fullstrong setup to use to have a fair baseline for my own (learned) branching rule. In particular, I'm unsure how to set the "idempotent" parameter of Vanillafullstrong branching as I do not fully understand what it does. I read the documentation and looked into the source code (https://scipopt.org/doc/html/branch__vanillafullstrong_8c_source.php) and, apparently, the idempotent parameter controls whether pseudo costs are updated (but these are not relevant for vanillafullstrong branching, are they?) and whether lower bounds on the child nodes get updated. How does the latter affect B&B tree sizes? I would have expected that the Strong Branching "side effects" mainly affect runtimes but not so much tree sizes but as soon as I set idempotent=True, tree sizes increase significantly. Here are some exemplary node averages when I solve 10 setcovering instances with three random seeds, using different branching rules: {relpscost: 330.67, fullstrong: 44.33, vanillafullstrong (idempotent=False): 103.63, vanillafullstrong (idempotent=True): 299.57}. This is just a small test set with small instances but I observed the same behavior on other problem classes and instance sizes. Why are the B&B trees so much larger when setting idempotent=True and what is the correct vanilla setup to have a fair comparison?


Thanks in advance!


Kind regards,

Caroline Spieckermann


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listserv.zib.de/pipermail/scip/attachments/20230816/dc25b951/attachment.html>