让我们首先将您的代码包装在一个函数中。我还添加了set.seed
命令以使结果可重现。您需要在运行模拟之前删除它们。
sim1 <- function(reps=50, steps=100 ) {
N<-5 # number of sites
sites<-LETTERS[seq(from=1,to=N)]
to.r<-rbind(sites)
p.move.r<-seq.int(0.05,0.90,by=0.05) # prob of moving to a new site
p.leave<-0.01*p.move.r # prob of leaving the system w/out returning
p.move.out<-0.01*p.move.r # prob of moving in/out
p.stay<-1-(p.move.r+p.leave+p.move.out) # prob of staying in the same site
set.seed(42)
random<-runif(10000,0,1) # generating numbers from a random distribution
cumsum.move <- read.table(text="A B C D E NA. left
A 0.0820000 0.3407822 0.6392209 0.3516242 0.3925942 0.1964 0.1964
B 0.1254937 0.4227822 0.6940040 0.3883348 0.4196630 0.3928 0.3928
C 0.7959865 0.8730183 0.7760040 0.7930623 0.8765180 0.5892 0.5892
D 0.8265574 0.8980259 0.8095507 0.8750623 0.9000000 0.7856 0.7856
E 0.9820000 0.9820000 0.9820000 0.9820000 0.9820000 0.9820 0.9820
NA. 0.9910000 0.9910000 0.9910000 0.9910000 0.9910000 0.9910 0.9910
left 1.0000000 1.0000000 1.0000000 1.0000000 1.0000000 1.0000 1.0000",header=TRUE)
cumsum.move <- as.matrix(cumsum.move)
for(o in 1:reps){
result<-matrix(as.character(""),steps) # Vector for storing sites
set.seed(42)
x<-sample(random,steps,replace=TRUE) # sample array of random number
time.step<-data.frame(x) # time steps used in the simulation (i)
colnames(time.step)<-c("time.step")
time.step$event<-""
set.seed(41)
j<-sample(1:N,1,replace=T) # first column to be selected
set.seed(40)
k<-sample(1:N,1,replace=T) # selection of column for ind. that move in/out
for(i in 1:steps){
for (t in 1:(N+1)){
if(time.step$time.step[i]<cumsum.move[t,j]){
time.step$event[i]<-to.r[t]
break
}
}
ifelse(time.step$event[i]=="",break,NA)
result[i]<-time.step$event[i]
j<-which(to.r==result[i])
if(length(j)==0){j<-k}
}
result<-time.step$event
}
result
}
请注意,result
在 o 的每次迭代中都会被覆盖。我不认为你想要那个,所以我解决了这个问题。此外,您使用data.frame
循环内部。作为一般规则,您应该避免data.frames
像瘟疫这样的内部循环。尽管它们非常方便,但就效率而言,它们很糟糕。
sim2 <- function(reps=50, steps=100) {
N<-5 # number of sites
sites<-LETTERS[seq(from=1,to=N)]
to.r<-rbind(sites)
p.move.r<-seq.int(0.05,0.90,by=0.05) # prob of moving to a new site
p.leave<-0.01*p.move.r # prob of leaving the system w/out returning
p.move.out<-0.01*p.move.r # prob of moving in/out
p.stay<-1-(p.move.r+p.leave+p.move.out) # prob of staying in the same site
set.seed(42)
random<-runif(10000,0,1) # generating numbers from a random distribution
cumsum.move <- read.table(text="A B C D E NA. left
A 0.0820000 0.3407822 0.6392209 0.3516242 0.3925942 0.1964 0.1964
B 0.1254937 0.4227822 0.6940040 0.3883348 0.4196630 0.3928 0.3928
C 0.7959865 0.8730183 0.7760040 0.7930623 0.8765180 0.5892 0.5892
D 0.8265574 0.8980259 0.8095507 0.8750623 0.9000000 0.7856 0.7856
E 0.9820000 0.9820000 0.9820000 0.9820000 0.9820000 0.9820 0.9820
NA. 0.9910000 0.9910000 0.9910000 0.9910000 0.9910000 0.9910 0.9910
left 1.0000000 1.0000000 1.0000000 1.0000000 1.0000000 1.0000 1.0000",header=TRUE)
cumsum.move <- as.matrix(cumsum.move)
res <- list()
for(o in 1:reps){
result<-character(steps) # Vector for storing sites
set.seed(42)
time.step<-sample(random,steps,replace=TRUE) # sample array of random number
#time.step<-data.frame(x) # time steps used in the simulation (i)
#colnames(time.step)<-c("time.step")
#time.step$event<-""
event <- character(steps)
set.seed(41)
j<-sample(1:N,1,replace=T) # first column to be selected
set.seed(40)
k<-sample(1:N,1,replace=T) # selection of column for ind. that move in/out
for(i in 1:steps){
for (t in 1:(N+1)){
if(time.step[i]<cumsum.move[t,j]){
event[i]<-to.r[t]
break
}
}
ifelse(event[i]=="",break,NA)
result[i]<-event[i]
j<-which(to.r==result[i])
if(length(j)==0){j<-k}
}
res[[o]]<-event
}
do.call("rbind",res)
}
两个函数给出相同的结果吗?
res1 <- sim1()
res2 <- sim2()
all.equal(res1,res2[1,])
[1] TRUE
新版本更快吗?
library(microbenchmark)
microbenchmark(sim1(),sim2())
Unit: milliseconds
expr min lq median uq max
1 sim1() 204.46339 206.58508 208.38035 212.93363 269.41693
2 sim2() 77.55247 78.39698 79.30539 81.73413 86.84398
嗯,三倍已经很不错了。我看不到进一步改进循环的可能性,因为那些break
s. 这只剩下并行化作为一种选择。
sim3 <- function(ncore=1,reps=50, steps=100) {
require(foreach)
require(doParallel)
N<-5 # number of sites
sites<-LETTERS[seq(from=1,to=N)]
to.r<-rbind(sites)
p.move.r<-seq.int(0.05,0.90,by=0.05) # prob of moving to a new site
p.leave<-0.01*p.move.r # prob of leaving the system w/out returning
p.move.out<-0.01*p.move.r # prob of moving in/out
p.stay<-1-(p.move.r+p.leave+p.move.out) # prob of staying in the same site
set.seed(42)
random<-runif(10000,0,1) # generating numbers from a random distribution
cumsum.move <- read.table(text="A B C D E NA. left
A 0.0820000 0.3407822 0.6392209 0.3516242 0.3925942 0.1964 0.1964
B 0.1254937 0.4227822 0.6940040 0.3883348 0.4196630 0.3928 0.3928
C 0.7959865 0.8730183 0.7760040 0.7930623 0.8765180 0.5892 0.5892
D 0.8265574 0.8980259 0.8095507 0.8750623 0.9000000 0.7856 0.7856
E 0.9820000 0.9820000 0.9820000 0.9820000 0.9820000 0.9820 0.9820
NA. 0.9910000 0.9910000 0.9910000 0.9910000 0.9910000 0.9910 0.9910
left 1.0000000 1.0000000 1.0000000 1.0000000 1.0000000 1.0000 1.0000",header=TRUE)
cumsum.move <- as.matrix(cumsum.move)
#res <- list()
#for(o in 1:reps){
cl <- makeCluster(ncore)
registerDoParallel(cl)
res <- foreach(1:reps) %dopar% {
result<-character(steps) # Vector for storing sites
set.seed(42)
time.step<-sample(random,steps,replace=TRUE) # sample array of random number
#time.step<-data.frame(x) # time steps used in the simulation (i)
#colnames(time.step)<-c("time.step")
#time.step$event<-""
event <- character(steps)
set.seed(41)
j<-sample(1:N,1,replace=T) # first column to be selected
set.seed(40)
k<-sample(1:N,1,replace=T) # selection of column for ind. that move in/out
for(i in 1:steps){
for (t in 1:(N+1)){
if(time.step[i]<cumsum.move[t,j]){
event[i]<-to.r[t]
break
}
}
ifelse(event[i]=="",break,NA)
result[i]<-event[i]
j<-which(to.r==result[i])
if(length(j)==0){j<-k}
}
#res[[o]]<-event
event
}
stopCluster(cl)
do.call("rbind",res)
}
结果一样吗?
res3 <- sim3()
all.equal(res1,c(res3[1,]))
[1] TRUE
快点?(让我们在我的 Mac 上使用 4 个内核。您可能会尝试访问具有更多内核的服务器。)
microbenchmark(sim1(),sim2(),sim3(4))
Unit: milliseconds
expr min lq median uq max
1 sim1() 202.28200 207.64932 210.32582 212.69869 255.2732
2 sim2() 75.39295 78.95882 80.01607 81.49027 125.0866
3 sim3(4) 1031.02755 1046.41610 1052.72710 1061.74057 1091.2175
那看起来很可怕。但是,该测试对并行功能是不公平的。该函数被调用 100 次,只有 50 次重复。这意味着我们获得了并行化的所有开销,但几乎没有从中受益。让我们更公平:
microbenchmark(sim1(rep=10000),sim2(rep=10000),sim3(ncore=4,rep=10000),times=1)
Unit: seconds
expr min lq median uq max
1 sim1(rep = 10000) 42.16821 42.16821 42.16821 42.16821 42.16821
2 sim2(rep = 10000) 16.13822 16.13822 16.13822 16.13822 16.13822
3 sim3(ncore = 4, rep = 10000) 38.18873 38.18873 38.18873 38.18873 38.18873
更好,但仍然不令人印象深刻。如果复制和步骤的数量进一步增加,并行功能看起来不错,但我不知道你是否需要。