In my stochastic processes class, Prof Mike Steele assigned a homework problem to calculate the ruin probabilities for playing a game where you with 1 dollar with probability p and lose 1 dollar with probability 1-p. The probability of winning is not specified, so it can be a biased game. Ruin probabilities are defined to be the probability that in a game you win 10 before losing 10, win 25 before losing 25, and win 50 before losing 50, etc. In total, I found three distinct methods to calculate.

This is a particularly great example to illustrate how to solve a problem using three fundamentally different methods: the first is theoretical calculation, second is simulation to obtain asymptotic values, and third is numerical linear algebra (matrix algorithm) which also gives exact values.


Method 1: First Step Analysis and Direct Computation of Ruin Probabilities

Let h(x) be the probability of winning $n before losing stake of x dollars.

First step analysis gives us a system of three equations: h(0) = 0; h(n) = 1; h(x) = p*h(x+1) + (1-p)*h(x-1).

How to solve this system of equations? We need the "one" trick and the telescoping sequence.

The trick is: (p + (1-p)) * h(x) = h(x) = p*h(x+1) + (1-p)*h(x-1) => p*(h(x+1) - h(x)) = (1-p)*(h(x) - h(x-1)) => h(x+1) - h(x) = (1-p)/p * (h(x)-h(x-1))

Denote h(1) - h(0) = c, which is unknown yet, we have a telescoping sequence: h(1) - h(0) = c; h(2) - h(1) = (1-p)/p * c; h(3) - h(2) = ((1-p)/p)^2 * c ... h(n) - h(n-1) = ((1-p)/p)^(n-1) * c.

Now, add up the telescoping sequence and use the initial conditions, we get: 1 = h(n) = c*(1+ ((1-p)/p) + ((1-p)/p)^2 + ... + ((1-p)/p)^(n-1)) => c = (1 - (1-p)/p) / (1 - ((1-p)/p)^N-1). So h(x) = c * (((1-p)/p) ^ x - 1) / ((1-p)/p)-1) = (((1-p)/p) ^ x - 1) / (((1-p)/p)^N - 1)


Method 2: Monte Carlo Simulation of Ruin Probabilities

The idea is to simulate sample paths from initial stake of x dollars and stop when it either hits 0 or targeted wealth of n.

We can specify the number of trials and get the percentage of trials which eventually hit 0 and which eventuallyhit n. This is important - in fact, I think the essence of Monte Carlo method is to have a huge number of trials to maintain accuracy, and to get a percentage of the number of successful trials in the total number of trials.

In each step of a trial, we need a Bernoulli random variable (as in a coin flip) to increment x by 1 with probability p and -1 with probability 1-p.

In Python this becomes:

from numpy import random
import numpy as np def MC(x,a,p):
  end_wealth = a
  init_wealth = x
  list = []
  for k in range(0, 1000000):
    while x!= end_wealth and x!= 0:
      if np.random.binomial(1,p,1) == 1:
        x += 1
      else:
        x -= 1
    if x == a:
      list.append(1)
    else:
      list.append(0)
  x = init_wealth
  print float(sum(list))/len(list) MC(10,20,0.4932)
MC(25,50,0.4932)
MC(50,100,0.4932)

You can see the result of this simulation by plugging in p = 0.4932 = (18/37)*.5 + .5*.5 = 0.4932, which is the probability of winning the European Roulette with prisoner's rule. As the number of trials get bigger and bigger, the result gets closer and closer to the theoretical value calculated under Method 1.


Method 3: Tridiagonal System

According to wiki, a tridiagonal system has the form of a_i * x_i-1 + b_i * x_i + c_i * x_i+1 = d_i where i's are indices.

It is clear that the ruin problem exactly satisfies this form, i.e.  h(x) := probability of winning n starting from i, h(x) = (1-p)*h(x-1) + p*h(x+1) => -(1-p)*h(x-1) + h(x) -p*h(x+1) = 0, h(0) = 0, h(n) = 1.

And therefore, for the tridiagonal matrix, the main diagonal consists of 1's, and the upper diagonal consists of -(1-p)'s, and the lower diagonal consists of -p's.

In Python this becomes:

import numpy as np
from scipy import sparse
from scipy.sparse.linalg import spsolve n = 100
p = 0.4932
q = 1-p d_main = np.ones(n+1)
d_super = -p * d_main
d_super[1] = 0
d_sub = -q * d_main
d_sub[n-1] = 0 data = [d_sub, d_main, d_super]
print data
A = sparse.spdiags(data, [-1,0,1], n+1, n+1, format='csc') b = np.zeros(n+1)
b[n] = 1
x = spsolve(A, b)
print x

Gambler's Ruin Problem and 3 Solutions的更多相关文章

  1. [Introduction to programming in Java 笔记] 1.3.8 Gambler's ruin simulation 赌徒破产模拟

    赌徒赢得机会有多大? public class Gambler { public static void main(String[] args) { // Run T experiments that ...

  2. 比特币_Bitcoin 简介

    2008-11   Satoshi Nakamoto  Bitcoin: A Peer-to-Peer Electronic Cash System http://p2pbucks.com/?p=99 ...

  3. Bitcoin: A Peer-to-Peer Electronic Cash System

    Bitcoin: A Peer-to-Peer Electronic Cash System Satoshi Nakamoto October 31, 2008 Abstract A purely p ...

  4. Mathematics for Computer Science (Eric Lehman / F Thomson Leighton / Albert R Meyer 著)

    I Proofs1 What is a Proof?2 The Well Ordering Principle3 Logical Formulas4 Mathematical Data Types5 ...

  5. [0x01 用Python讲解数据结构与算法] 关于数据结构和算法还有编程

    忍耐和坚持虽是痛苦的事情,但却能渐渐地为你带来好处. ——奥维德 一.学习目标 · 回顾在计算机科学.编程和问题解决过程中的基本知识: · 理解“抽象”在问题解决过程中的重要作用: · 理解并实现抽象 ...

  6. URAL 1430 Crime and Punishment

    Crime and Punishment Time Limit:500MS     Memory Limit:65536KB     64bit IO Format:%I64d & %I64u ...

  7. Attention and Augmented Recurrent Neural Networks

    Attention and Augmented Recurrent Neural Networks CHRIS OLAHGoogle Brain SHAN CARTERGoogle Brain Sep ...

  8. Win7 服务优化个人单机版

    我的PC设备比较旧了,为了系统能流畅点,不必要的服务就不开启了.然而,服务那么多,每次重装,都要从头了解一下一边,浪费时间. 个人在网络上收集信息并结合自己的摸索,整理如下,以备查找. 服务名称  显 ...

  9. [转]WIN7服务一些优化方法

    本文转自:http://bbs.cfanclub.net/thread-391985-1-1.html Win7的服务,手动的一般不用管他,有些自动启动的,但对于有些用户来说是完全没用的,可以考虑禁用 ...

随机推荐

  1. matlab图

    .6 统计作图 4.6.1 正整数的频率表 命令 正整数的频率表 函数 tabulate 格式 table = tabulate(X) %X为正整数构成的向量,返回3列:第1列中包含X的值第2列为这些 ...

  2. 数据库imp导表dmp的方法

    1>sqlplus / as sysdba 进入sqlplus 2>drop user USER cascade 3>create user USER IDENTIFIED BY P ...

  3. spring mvc 第一天【注解实现springmvc的基本配置】

    创建pojo,添加标识类的注解@Controller,亦可添加该Handler的命名空间:设置类的@RequestMapping(value="/hr") 该类中的方法(Handl ...

  4. 简例 一次执行多条mysql insert语句

    package com.demo.kafka;import java.sql.Connection;import java.sql.DriverManager;import java.sql.Prep ...

  5. git 创建远程分支和删除 master 分支

    . . . . . 最近需要将不同的客户的代码分开管理,所以需要为这些代码分别创建分支. 目前版本库中分支结构如下: [yuhuashi@local:Project]$ git branch -a* ...

  6. 史上最全的Python电子书教程资源下载(转)

    网上搜集的,点击即可下载,希望提供给有需要的人^_^   O'Reilly.Python.And.XML.pdf 2.02 MB   OReilly - Programming Python 2nd. ...

  7. C# WinForm 中 MessageBox的使用详解

    1.C# WinForm 中 MessageBox的使用详解:http://www.cnblogs.com/bq-blog/archive/2012/07/27/2611810.html

  8. Python 9X9乘法口诀表

    #乘法口诀 for i in range(1,10): for j in range(1,i+1): print ( j,'*',i,'=',j*i,'',' ',end='') print(end= ...

  9. JavaScript 常用函数总结

    javascript函数:  ·常规函数  ·数组函数  ·日期函数  ·数学函数  ·字符串函数 .cookie函数 1.常规函数 javascript常规函数包括以下9个函数:  (1)alert ...

  10. Centos版Linux 一些常用操作命令

    Linux命令收集 1.文件处理命令:ls 功能描述:显示目录文件 命令英文原意:list 命令所在路径:/bin/ls 执行权限:所有用户 语法:  ls  选项[-ald]  [文件或目录] -a ...