Jean-Francois Baffier
Jean-Francois Baffier
Not only bad variables should be selected but occasionally some good variables too to avoid local extrema. @richoux I mention you here as it is also related to renewal of...
Two candidates: - [ ] uniform weights (so only the distance to the selected variable) - [ ] arbitrary weights through multiplex networks
The candidate method is Roulette Selection. The goal is to reduce the computation time required to select a bad [enough] variable (instead of the worst).
Examples: - [ ] Physical object (in Garamon.jl) - [ ] sub problems as multidimensional variable - [ ] a LP/MILP/etc. compatible subproblem
Requires: - [ ] Asynchronous solving - [ ] Pool of solution - [ ] Restart strategy - [ ] Dynamic re-computation (incrementally) of preprocesses (such as Neighborhood)