max - Solving stochastic maximum bipartite matching problem -
i have faced following problem:
- there 2 disjoint sets,
a
,b
- for each pair of elements (
a
,b
) (a
belongs seta
,b
belongs setb
) there probabilitypij
known in advance. represents probability (certainty level)a
matchesb
, or in other words, how closelya
matchesb
(and vice-versa, becausepij
==pji
). - i have find matching highest probability/certainty , find out pairs (
a
,b
) describe matching - every element must matched / paired other set once (like in standard bipartite matching problem)
- if possible, compute number approximately expresses uncertainty level obtained matching (let's 0 represents random guess , 1 represents certainty)
a simple practical example in such algorithm required described below (this not problem solving!):
- two people asked write letters - z on piece of paper
- for each pair of letters (
a
,b
) run pattern matcher determine probability lettera
written persona
represents letterb
wrote personb
. gives probability matrix expresses kind of similarity correlation each pair of letters (a
,b
) - for each letter person
a
wrote, need find corresponding letter written personb
current approach: wondering if assign weights proportional logarithm of certainty level / probability element a
set a
matches element b
set b
, run maximum weighted bipartite matching find maximum sum. logarithm because want maximize total probability of multiple matching, , since single matches (represented pairs of matched elements a
- b
) form chain of events, product of probabilities, taking logarithm converts sum of probabilities, maximized using algorithm weighted bipartite matching, such hungarian algorithm. somehow doubt approach ensure best matching in terms of statistical expected maximum.
after searching bit, closest problem found two-stage stochastic maximum weighted matching problem, np-hard, need kind of "one-stage" stochastic maximum weighted matching problem.
i wonder if can use maxflow/mincut. can't prove it's optimal @ moment, problem may np-hard anyway. can use mf/mc find perfect matching when have bipartite graph v=(a,b) creating source connected nodes in weight of 1 , sink connected nodes in b weight 1. i'm proposing make weights of edges cross b probabilities mentioned above. think?
Comments
Post a Comment