Well, to be honest I found few likeness.
Are your SDDP.jl scenario tree based, IIRC? Therefore you have the node concept.
Multistage stochastic programs (MSSP) are very hard to solve to optimality. (Actually my problem can be deemed a Two stage 1 scenario MIP with J blocks. My problem is already hard when J is large, let alone >2 stage multiple scenarios.)
Seems you’re using the Distributed module. I currently have no idea about it.
Your algorithm:
@sync for _ in 1:max_threads
Threads.@spawn begin
where max_threads is hardware based. By comparison, my algorithm is hardware-agnostic (as long as there are multiple logical processors). And you used @sync, which I hadn’t use it, neither wait the subproblem tasks. You made use of Channels, but I didn’t (as my comment in the #25 post).
One thing I’m most curious about: how can you run multiple forward-backward pass? Is this meaningful? i.e. Can the SDDP training algorithm be run in parallel?
Well, actually not that curious, since I’m not that interested on MSSP anymore. (Therefore you can opt to neglect my questions… (I’m somewhat not very spirited on this.))
Generally speaking, I think JuMP is more flexible than SDDP.jl (or some other extension packages). I find it relatively easy to write algorithms with JuMP—although it as well takes me years to get familiar with it. I was a noob with little experience then. But happily I’ve gained quite a few proficiency now.
I’m unsure if I’m the first to try writing such an asynchronous cutting plane algorithm.
I’m also okay being the second.