Solving Multilocal Optimization Problems with a Recursive Parallel Search of the Feasible Region

Abstract

Abstract. Stretched Simulated Annealing (SSA) combines simulated annealing with a stretching function technique, in order to solve multilocal programming problems. This work explores an approach to the parallelization of SSA, named PSSA-HeD, based on a recursive heterogeneous decomposition of the feasible region and the dynamic distribution of the resulting subdomains by the processors involved. Three PSSAHeD variants were implemented and evaluated, with distinct limits on the recursive search depth, offering different levels of numerical and computational efficiency. Numerical results are presented and discussed.

coPSSA - Constrained Parallel Stretched Simulated Annealing

“…The Stretched Simulated Annealing algorithm stops when no new optimum is identified after r consecutive runs. For more details see [14], [15].…”

“…PSSA executions took place in the 8 worker nodes; these offer a total of 32 CPU-cores, fully specified in the base MPI hostfile supplied to coPSSA, that are used, four at a time, to service the PSSA execution requests of each coPSSA task; thus, each PSSA execution always consumed 4 cores, with 1 core for the master process, and 3 cores for slave processes. In order to fully exploit the 3 slave cores, the number of subdomains processed by PSSA was defined to be no less (and as close as possible) than 3; for 2-dimensional problems, like the ones we tested, this is achieved with a granularity g = 0.5, that generates 4 sub-domains [15]. The PSSA variant used was always the HoD variant, once it is the fastest and uses a fixed number of sub-domains (4, in our evaluation scenario).…”

“…In previous work [15], [17], Parallel Stretched Simulated Annealing (PSSA) was presented (and successively refined) as a parallel version of SSA, based on the decomposition of the initial search domain in several sub-domains to which SSA is independently applied by a set of processors. Several domain decomposition and work assignment approaches were explored, ranging from homogeneous data decomposition and static work assignment, to heterogeneous data decomposition and dynamic work assignment, leading to successively increasing levels of numerical efficiency (here, the main contribution of parallelization is not to decrease optima search times, but to increase the number of optima found in a bounded time).…”

Solving Constrained Multilocal Optimization Problems with Parallel Stretched Simulated Annealing

“…The Stretched Simulated Annealing algorithm stops when no new optimum is identified after r consecutive runs. [15,16] provide more details.…”

“…In previous work [16,18], Parallel Stretched Simulated Annealing (PSSA) was introduced as a parallel version of SSA, based on the decomposition of the feasible region in several subregions to which SSA is independently applied by a set of processors. Several domain decomposition and distribution approaches were explored, leading to successively increasing levels of numerical efficiency.…”