Open odow opened 6 years ago
I've just realized that https://github.com/JuliaOpt/JuMP.jl/pull/1209 will help speed things up heaps when we swap between solving the MIPs and LP relaxation.
The biggest bottlenecks are a) Getting the methods that solve for the duals to converge b) The fact that we sometimes find duals that are not very good because of primal degeneracy
I tried @eyobzewdie's booking management problem with 6 days, 4 rooms (i.e. 24 states), and 10 stages on my workstation (running in parallel with 4 core), and after 12 hours it had only done 1000 iterations with the subgradient method and the bound hadn't moved from the intial bound. The level method did slightly better (535 iterations, and the bound dropped slightly).
It seems the current implementation just can't handle these problem sizes. Any idea's on how to speed up the solves? Or different methods for the lagrangian solver?