文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > Simulated Annealing for Graph Bisection
侵权投诉

Simulated Annealing for Graph Bisection

Simulated Annealing for Graph Bisection

Commencing with k= (n1=2), the bias f+? f?= (minfkn?2=3+"; 1g)= (n?1=6+") can naively be expected, over n2=3 moves, to contribute an imbalance of (n1=2+"). Over the same time, as the transition probabilities are all approximately 1, the number of 8 k-increasing moves has variance (n2=3) and is exponentially unlikely to di er by more than O(n1=2) from its nominal value; the same is true for k-decreasing moves; and so with exponential likelihood the nal imbalance k will be (n1=2+" ) O(n1=2)= (n1=2+" ). Another n2=3 moves starting from this point would by the same reasoning produce k= (n1=2+2" ), with (1=") repetitions yielding k= (n2=3). The total number of moves used, (1=") n2=3, is within the allotment t= O(n2=3+" ). Again, this merely provides intuition for how k grows: we will take it up again formally after showing that -atypical states can be avoided under the more rigorous parameter values that will be required.

5 An Improved Argument The previous section's proof that no atypical state is encountered could be thought of as an adversary argument: even an adversary with freedom to accept or reject each move cannot guide the process to a partition for which (G; ) is atypical. To allow extension of th

e result to longer annealing runs with more exacting parameter values, in this section we make a roughly analogous argument which capitalizes on the fact that the adversary actually has limited freedom. Speci cally, for almost all G 2 G4n;p;r, and every bisection, the transition probabilities from must\on average" lie within a small range about their expectation over G. If the adversary is con ned to choosing a probability within this range and then respecting a fair coin toss, there is little chance it can alter the outcome. Thus in t moves it makes far fewer than t decisions, allowing a larger t than previously. For a given graph G, let P (; 0)= PG (; 0) be the probability that from state the transition 0 is made (generated and accepted), and let P (; 0)= EG P (; 0). For a 0) given G and, let kP? P k= kP (; ); P (; )k= jP (; 0)? P (; pj. Call a pair (G; )\rough" if kP? P k exceeds some threshold, which we set to be O( n=T ). b b b P 0

b

b

b for all, kP? P k= O ( n=T ):

Claim 5.1 For any temperature T> 0, with overwhelming probability over G 2 G4n;p;r, p 6= . First,

Proof Let X 0 be the cost increase associated with a possible transition 7! 0, for 0 EG jP (; 0)? EG P (; 0)j= EG P (; 0)? EG P (; 0)]2 EG P (; 0)? EG P (; 0)]2 1= 4n2 EG A(X 0)? EG A(X 0)]2 q q q

(4)

上一页第9页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email