Download presentation
Presentation is loading. Please wait.
1
Monte Carlo Simulation Part.1 Dept. Phys., Tunghai Univ. Numerical Methods, C. T. Shih
2
如何以蒙地卡羅法做數值積分 If we want to find the numerical integration: We can simply divide the region [0,1] evenly into M slices and the integral can be approximated by:
3
Simple Sampling On the other hand, we can select x n for n=1,2,...,M from a uniform random number generator (亂數產生器) If M is large enough, x n is set of numbers uniformly distributed in the region [0,1]
4
Simple Sampling (conti.) The integral can be approximated by: Now the error in the integral comes from the fluctuation of the distribution of x n :
5
f(x) 01
6
Random Number Generator Indeed, there is no real random number generator in a computer program – the program always follows the commands by the programmer We can only have a “pseudo-random number generator” A good random number generator: very long period, good randomness (small correlation), fast
7
Pseudo-RNG Long period: what is the possible longest period on a computer? Answer: the largest integer, i.e., 2 31 -1 = 2,147,483,647 on a 32-bit computer
8
Linear Congruent Scheme The random numbers are generated by: a, b, c are “magic numbers” which determine the quality of the RNG One possible choice (tested OK): a = 7 5 = 16807, b = 0, and c = 2 31 -1 = 2,147,483,647 Also you need a “seed” of your RNG as the initial x
9
Procedure of the RNG 讀入系統時間(利用內建 TIME() 函數)作 為 ISEED ,也就是 x 0 利用 計算 x 1 將 x 1 作為新的 ISEED ,而所得之 random number 為 x 1 /c 依此類推,可得到一連串介於 (0,1) 之間 的均勻亂數
10
Overflow 問題 由於 可能發生 overflow ,因此程式上必須考慮:
11
練習: 利用 Monte Carlo 法計算以下函數的積分, M (取樣數為 100, 1000, 1000 ), a 取 0.2, 0.5, 1, 5, 10 ,求積分之近似值以及標 準差
12
Alternative: 一維函數的蒙地卡羅積分也可以用以下的方法 計算: 產生兩個亂數所成數對( x,y ), x 的範圍在積 分上下限 [a,b] 之間, y 的範圍在 [c,d] 之間, 且 c 不大於 f(x) 於 [a,b] 間之極小值, d 不小於 f(x) 於 [a,b] 間之極大值 若 0<y<f(x) 則 N + 增加一,若 f(x)<y<0 則 N - 增 加一,若總共產生 N 個亂數,則面積之平均值 為 :
14
平均值與標準差計算方式 上述過程每次產生 N 組亂數,可得到一 個 S 重複以上過程 M 次,即可得到 M 個 S i, i=1,2,...,M 求這 M 個 S 的平均值即為面積 標準差
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.