I have been curious for a while as to how R might be used for the construction of an individually-based model (IBM), or agent-based model (ABM). In particular, what R objects lend themselves best to storing information on individuals, and allow for new individuals to be added or subtracted throughout the simulation?

In this first attempt, I have ended up opting for a multi-level list, where elements represent individuals, and sub-levels contain attribute information. The main reason is speed - In a previous post I highlighted the fact that lists are not penalized in the same way as a data.frame when the object in "grown" or concatenated with additional information (due to time spent reallocating memory).

The example models a population with a given birth rate, death rate, and carrying capacity. Attributes of individuals that are recorded in the list include their age, whether they are alive or dead, and their color (blue or red). The attribute of color is passed from parent to offspring, and there is a tendency for one phenotype to dominate over time. The idea comes from a simple model of genetic drift that can be explored with the IBM programming platform NetLogo.

So far so good, but there are still some speed issues associated with a list that continues to grow. Part of the speed issue is due to the calculation of summary statistics during each iteration (using e.g. sapply). A cropping of the list to retain only alive individuals, has dramatic improvements on speed as well, so there appears to be a cost associated with growing the list object itself. The difference might be that my previous example was filling a

*empty*list, where the number of elements was predefined.

I would be interested if anyone has any thoughts on this or, more generally, on the construction of IBMs in R. While R is probably not the best programming language suitable to IBMs, it would be interesting to know if more examples exist.

**Example:**

b <- 0.14 # probability of birth d <- 0.08 # probability of death K <- 100 # carrying capacity N0 <- 50 # starting number of individuals t <- 500 # time of simulation #create starting individual w attributes ("alive", "age", "color") set.seed(1) ind <- vector(mode="list", N0) for(i in seq(ind)){ ind[[i]]$alive <- 1 ind[[i]]$age <- 0 ind[[i]]$color <- c("blue", "red")[round(runif(1)+1)] } #make empty vectors to record population statistics time <- seq(t+1) pop <- NaN * time # population size pop[1] <- N0 frac.blue <- NaN * time # fraction of population that is blue cols <- sapply(ind, function(x) x$color) frac.blue[1] <- sum(cols == "blue") / length(cols) med.age <- NaN * time ages <- sapply(ind, function(x) x$age) med.age[1] <- median(ages) #simulation save.alive.only <- TRUE # optional cropping of "ind" to include alive individuals only t1 <- Sys.time() for(i in seq(t)){ # loop for each time increment is.alive <- which(sapply(ind, function(x) x$alive) == 1) for(j in is.alive){ #loop for each alive individual birth <- runif(1) <= (b * (1 - length(is.alive)/K)) # calculate a birth probability for each individual that is alive if(birth){ len.ind <- length(ind) ind[[len.ind+1]] <- list(alive=1, age=0, color=ind[[j]]$color) # create offspring, inherits color of parent } death <- runif(1) <= d # calculate a death probability for each individual if(death){ ind[[j]]$alive <- 0 # if death, reset alive = 0 } else { #else, advance age + 1 ind[[j]]$age <- ind[[j]]$age + 1 # advance age of parent } } #optional cropping of list "ind" if(save.alive.only){ is.dead <- which(sapply(ind, function(x) x$alive) == 0) if(length(is.dead) > 0) ind <- ind[-is.dead] } #Population stats is.alive <- which(sapply(ind, function(x) x$alive) == 1) pop[i+1] <- length(is.alive) cols <- sapply(ind, function(x) x$color) frac.blue[i+1] <- sum(cols[is.alive] == "blue") / length(is.alive) ages <- sapply(ind, function(x) x$age) med.age[i+1] <- median(ages[is.alive]) print(paste(i, "of", t, "finished", "[", round(1/t*100), "%]")) } t2 <- Sys.time() dt <- t2-t1 dt #plot populations png("pops_vs_time.png", width=6, height=4, units="in", res=400) par(mar=c(4,4,1,1)) pop.blue <- pop * frac.blue pop.red <- pop * (1-frac.blue) ylim=range(c(pop.blue, pop.red)) plot(time, pop.blue, t="l", lwd=2, col=4, ylim=ylim, ylab="Population size") lines(time, pop.red, lwd=2, col=2) legend("topleft", legend=c("blue pop.", "red pop."), lwd=2, col=c(4,2), bty="n") dev.off() #plot median age png("med_age_vs_time.png", width=6, height=4, units="in", res=400) par(mar=c(4,4,1,1)) plot(time, med.age, t="l", lwd=2, ylab="Median age") dev.off()

I've been wondering about how to put together not just iBM's, but how to populate nicely-analyzable structures in general. What about putting individuals in an object-oriented framework, with each individual being represented by a single object? You would also have collections to access all individuals, etc.

ReplyDeleteI think this is the strategy of the R package simecol, but I haven't had time to look into it. Maybe you're willing to write a guest post on the approach?

DeleteV interesting post as I'd been wondering how to implement ABM in R. One thought - don't programs such as NetLogo act on individual living agents within a population at random? So would it be better to replace "for (j in is.alive)" with "for (j in sample(is.alive))" in your code?

ReplyDeleteI don't believe that that is a generalization that one has for all ABMs. You can have all sorts of interactions between agents - random, within a given distance, etc. In this very simple example, there is no interaction at all - only a chance of inheriting a given trait given the proportions of that trait in the population.

DeleteMarc,

ReplyDeleteThanks for your interesting post. I made some editions to your code: I got rid of some loops, I vectorized some computations. The code runs faster.

b = 0.14 # probability of birth

d = 0.08 # probability of death

K = 100 # carrying capacity

N0 = 50 # starting number of individuals

t = 500 # time of simulation

#create starting individual w attributes ("alive", "age", "color")

set.seed(1)

myind=cbind(alive=rep(1,N0),age=rep(0,N0),color=round(runif(N0)+1))

# color=c("blue", "red")[myind[,3]]

#make empty vectors to record population statistics

time = seq(t+1)

mypop = rep(NaN,t+1) # population size

mypop[1] = N0

myfrac.blue = rep(NaN,t+1) # fraction of population that4 is blue

myfrac.blue[1] = sum(myind[,'color'] == 1) / sum(myind[,1])

mymed.age = rep(NaN,t+1)

myages = myind[,'age']

mymed.age[1] = median(ages)

#simulation

save.alive.only = TRUE # optional cropping of "ind" to include alive individuals only

t1 = Sys.time()

for(i in 1:t){ # loop for each time increment

mybirths = rbind(c(0,0,0), myind[myind[,'alive']==1,])

n.alive=dim(mybirths)[1] - 1

if(n.alive){

# this is to reproduce the order of random variables in your code

ru=runif(2*n.alive)

# you use one call to runif for births and the next for deaths

#select births

mybirths[,1]=c(0, as.numeric(ru[2*(1:n.alive)-1]<=b*(1-n.alive/K)) )

mybirths[,2]=0

mybirths=mybirths[mybirths[,1]==1,]

#select survivors

myind[myind[,'alive']==1,1]=myind[myind[,'alive']==1,1]*(ru[2*(1:n.alive)]>d)

#advance age of survivors

myind[,2]=myind[,2]+myind[,1]

#optional cropping of list "ind"

if(save.alive.only) myind = myind[myind[,1]==1,]

#old & new

myind=rbind(myind,mybirths)

}

#Population stats

mypop[i+1] = sum(myind[,1])

myfrac.blue[i+1] = sum(myind[,'color']==1) / mypop[i+1]

mymed.age[i+1] = median(myind[myind[,1]==1,2])

print(paste(i, "of", t, "finished", "[", round(1/t*100), "%]"))

}

t2 <- Sys.time()

dt <- t2-t1

dt

#plot populations

#png("pops_vs_time.png", width=6, height=4, units="in", res=400)

par(mar=c(4,4,1,1))

mypop.blue <- mypop * myfrac.blue

mypop.red <- mypop * (1-myfrac.blue)

ylim=range(c(mypop.blue, mypop.red))

plot(time, mypop.blue, t="l", lwd=2, col=4, ylim=ylim, ylab="Population size")

lines(time, mypop.red, lwd=2, col=2)

legend("topleft", legend=c("blue pop.", "red pop."), lwd=2, col=c(4,2), bty="n")

#dev.off()

#plot median age

#png("med_age_vs_time.png", width=6, height=4, units="in", res=400)

par(mar=c(4,4,1,1))

plot(time, med.age, t="l", lwd=2, ylab="Median age")

#dev.off()

Thanks Victor! It is indeed much faster.

Delete