NP-hard resource allocation problems are best solved via ap- proximation techniques which produce acceptable solutions in the time available. The best of such "good-enough/soon- enough" techniques handle large problems, can run distributed, can adapt to changes to the problem while they are solving it, ensure that solution quality always increases, and start proposing solutions rapidly so that the user can quit opti- mizing at any time. And, maybe most importantly, they must let the user know when it makes sense to quit and when it is worthwhile to wait. Our Dynamic Marble Size (DMS) scheme appears promis- ing as an approach offering all those traits. It is a market- inspired distributed multi-agent scheme, in which coopera- tive task agents compete for resources through an auction but are capable of altruistically withdrawing to balance local vs. global concerns, reduce contention, and speed closure. In DMS, tasks maximally bid their value which maximizes other bidders' information about their prospects of succeed- ing, and thereby also guide them focus on resources they can win. An oscillation-avoiding bid adjustment algorithm uti- lizes binary search techniques to prevent those adjustments from introducing cycles or deadlocks. A "stubbornness- detection monitor" (a re-start limit) limits how many sets of resources a task will pursue. These three techniques com- bine to promote rapid convergence on good solutions - bid- ders avoid highly-contended resources and focus upon more promising alternatives. In addition, DMS schemes exhibit well-differentiated faster primary and slower secondary solution phases. They can de- tect which phase they are in, and thus guide the user in when it makes sense to stop and when it does not. Evaluations of DMS analyzed characteristics emerging from running very large numbers of randomly-generated prob- lems. In these evaluations we compare the solution qual- ity of DMS against results obtained with a Pseudo-Boolean encoding of the problem as the gold standard for central- ized solutions. For the types of problems that our problem
分享
引用
