I've implemented a simple dynamic programming example described here, using data.table, in the hope that it would be as fast as vectorized code.
library(data.table)
B=100; M=50; alpha=0.5; beta=0.9;
n = B + M + 1
m = M + 1
u <- function(c)c^alpha
dt <- data.table(s = 0:(B+M))[, .(a = 0:min(s, M)), s] # State Space and corresponging Action Space
dt[, u := (s-a)^alpha,] # rewards r(s, a)
dt <- dt[, .(s_next = a:(a+B), u = u), .(s, a)] # all possible (s') for each (s, a)
dt[, p := 1/(B+1), s] # transition probs
# s a s_next u p
# 1: 0 0 0 0 0.009901
# 2: 0 0 1 0 0.009901
# 3: 0 0 2 0 0.009901
# 4: 0 0 3 0 0.009901
# 5: 0 0 4 0 0.009901
# ---
#649022: 150 50 146 10 0.009901
#649023: 150 50 147 10 0.009901
#649024: 150 50 148 10 0.009901
#649025: 150 50 149 10 0.009901
#649026: 150 50 150 10 0.009901
To give a little content to my question: conditional on s
and a
, future values of s
(s_next
) is realized as one of a:(a+10)
, each with probability p=1/(B + 1)
. u
column gives the u(s, a)
for each combination (s, a)
.
- Given the initial values
V
(alwaysn by 1
vector) for each unique states
,V
updates according toV[s]=max(u(s, a)) + beta* sum(p*v(s_next))
(Bellman Equation). - Maximization is wrt
a
, hence,[, `:=`(v = max(v), i = s_next[which.max(v)]), by = .(s)]
in the iteration below.
Actually there is very efficient vectorized solution. I thought data.table
solution would be comparable in performance as vectorized approach.
I know that the main culprit is dt[, v := V[s_next + 1]]
. Alas, I have no idea how to fix it.
# Iteration starts here
system.time({
V <- rep(0, n) # initial guess for Value function
i <- 1
tol <- 1
while(tol > 0.0001){
dt[, v := V[s_next + 1]]
dt[, v := u + beta * sum(p*v), by = .(s, a)
][, `:=`(v = max(v), i = s_next[which.max(v)]), by = .(s)] # Iteration
dt1 <- dt[, .(v[1L], i[1L]), by = s]
Vnew <- dt1$V1
sig <- dt1$V2
tol <- max(abs(V - Vnew))
V <- Vnew
i <- i + 1
}
})
# user system elapsed
# 5.81 0.40 6.25
To my dismay, the data.table
solution is even slower than the following highly non-vectorized solution. As a sloppy data.table-user, I must be missing some data.table
functionality. Is there a way to improve things, or, data.table
is not suitable for these kinds of computations?
S <- 0:(n-1) # StateSpace
VFI <- function(V){
out <- rep(0, length(V))
for(s in S){
x <- -Inf
for(a in 0:min(s, M)){
s_next <- a:(a+B) # (s')
x <- max(x, u(s-a) + beta * sum(V[s_next + 1]/(B+1)))
}
out[s+1] <- x
}
out
}
system.time({
V <- rep(0, n) # initial guess for Value function
i <- 1
tol <- 1
while(tol > 0.0001){
Vnew <- VFI(V)
tol <- max(abs(V - Vnew))
V <- Vnew
i <- i + 1
}
})
# user system elapsed
# 3.81 0.00 3.81