## Combinatorial problem

Suppose you write the a number $$N$$ as the sum of $$k$$ non-negative terms. That is, $N = n_1 + n_2 + \cdots + n_k$ where $$n_i \ge 0$$ for $$i = 1,2,\ldots,k$$. For example, for $$k=4$$, $$N=10$$ can be written as $$N = 2 + 3 + 4 + 1$$, or as $$N = 5 + 0 + 2 + 3$$.

Here’s a challenge: Find the sum that maximizes a function $$f(n_1,n_2,\dots,n_k)$$. For example, let $$f(n_1, n_2, \dots, n_k) = (n_1 \times n_2 \times \cdots \times n_k)^\frac{1}{k}.$$ Plenty of other examples abound, but let’s use this simple function for now.

## R Implementation

Start by setting up the scene.

N <- 100
k <- 10

A state will refer to 10 numbers that sum to 100. Here is one way to obtain a random state:

vals <- sample(1:k, N, rep=T)
vals
##     6  4  9  3  4  9  7  6  4  8  9  3  1  4  3 10  5  9  3  7  3  7  3
##    7  3  5  7  5  5  8  5  4  4  7  7  1  7  1  8  4  3  2  3  8  5  1
##    3  1  5  2  2  7 10  6  7 10  9  4 10  5  6  4 10  7  4  1  1  9  1
##    7  9  8  5 10  9  2  3  8  9  1  7  6  7  6  7  9  1  2 10  1  5  1
##    7  6 10  3  7  1  3 10
state <- as.vector(table(vals))
state
##   13  5 13 10 10  7 17  6 10  9

Convince yourself that this really does produce a random state.

Now we want to maximize the product of any 10 values that sum to 100. Let’s slightly generalize this. We will maximize the log of the product (with a small tweak to account for a product that is 0). Clearly, finding $$(n_1, n_2, \ldots, n_k)$$ which maximizes the log of the product suffices.

f <- function(state){
##add a small value to get away from 0
-log(prod(state) + .001)/k
}

Did you notice the negative sign? Since we will use simulated annealing, we really want to solve a minimization problem. Thus, the minimum value of $$f$$ corresponds to the maximum product of the 10 numbers that sum to 100.

Try it out:

f(state)
##  -2.24152

## Proposal transition

The state spaces consists of all $$k-$$tuples of non-negative values that sum to $$N$$. If $$(n_1, n_2, \ldots, n_k)$$ is the current state, then we can propose to transition to a new state by:

• Randomly pick two indexes $$i,j \in \{1, 2, \ldots, k\}$$.
• If $$n_i > 0$$, then let
• $$n_i = n_i - 1$$
• $$n_j = n_j + 1$$
• If $$n_i = 0$$, then do nothing.

For example, suppose $$N = 10$$ and $$k = 4$$, and $$(0,2,5,3)$$ is the current state. Randomly select, say, $$i=2$$ and $$j=4$$. In this case, the new state becomes $$(0,1,5,4)$$. However, if we select $$i=1$$ and $$j=3$$, then (since $$n_1 = 0$$), the next state is the same as the current state.

Here is one way to do this in R:

state
##   13  5 13 10 10  7 17  6 10  9
sites <- sample(1:k, 2, rep=F)
sites
##  7 5
if(state[sites] > 0){
state[sites] <- state[sites] - 1
state[sites] <- state[sites] + 1
}
state
##   13  5 13 10 11  7 16  6 10  9

From this, create the proposal function.

proposal <- function(currState){
propState <- currState
sites <- sample(1:k, 2, rep=F)
if(propState[sites] > 0){
propState[sites] <- propState[sites] - 1
propState[sites] <- propState[sites] + 1
}
propState
}

Now create the doMove function. Notice how much it looks like our previous doMove function.

sig <- 0.1
doMove <- function(currState,sig){
# propose a transition
propState <- proposal(currState)
# get current and proposed function values
currF <- f(currState)
propF <- f(propState)
# calculate rho and make a move based on result
dFunc <- propF - currF
rho <- exp(-dFunc/sig)
if(runif(1,0,1) < rho){
return(propState)
}
return(currState)
}

Never hurts to test it…

state
##   13  5 13 10 11  7 16  6 10  9
(state <- doMove(state,1))
##   14  5 13  9 11  7 16  6 10  9
(state <- doMove(state,1))
##   14  5 13  9 11  7 16  6 11  8

## Simulated annealing

Here we go with simulated annealing. Let’s start clean.

N <- 100
k <- 10
(state <- as.vector(table(sample(1:k, N, rep=T))))
##   14  5 10 11  7 10  8 13 15  7

Ready to go. When doing simulated annealing, the choice of M and decFac are control parameters and needed to be fiddled with to get this to work properly.

sig <- .1
M <- 30000
decFac <- .9999
for(m in 1:M){
state <- doMove(state, sig)
# slowly decrease sig
sig <- sig*decFac
}
sig
##  0.00497796
state
##   13  9  7  8 12 10  9 12  9 11

As it turns out, the the answer is well-known—this function is maximized when all the values are as close to the same as possible.

Your turn: Change N and k, and see how the choices of $$\sigma^2$$, decFac, and M need to be adjusted to get a reasonable answer.

## Exercise: Different $$f(x)$$

There are lots of other interesting functions to consider. For example, try: $f(n_1, n_2, \dots, n_k) = \frac{1}{k}\sum_{i=1}^k n_i^2$ That is, the mean of the sum of the squares of $$(n_1, n_2, \dots, n_k)$$. There isn’t too much different here, but you will need to think about the initial value of $$\sigma^2$$ and how to decrease it properly.

## Assignment: Magic Squares

A $$3 \times 3$$ magic square is an arrangement of the numbers $$1, 2, \ldots, 9$$ in a $$3 \times 3$$ grid in such way that all the rows, columns, and diagonals have the same sum. Note there are 3 rows, 3 columns, and 2 diagonals—hence 8 sums total. For example:

$\begin{array}{|c|c|c|} \hline 4 & 9 & 2\\ \hline 3 & 5 & 7\\ \hline 8 & 1 & 6\\ \hline \end{array}$ Here, all the rows, columns, and diagonals sum to 15.

#### Find a magic square using simulated annealing!

• State space: An assignment of $$1, 2, \dots, 9$$ to the $$3 \times 3$$ grid.
• Function to minimize: Something which indicate how far the 8 sums corresponding to the three row, three column, and two diagonal sums are from the desired value of 15.
• Proposal transition: Do one of two things at each transitions.
• Pick two of the 9 sites and swap the values.
• Pick two rows or two columns and swap them

#### Starting state

Here is a simple way to create a random starting state. It almost certainly will not be magic!

mm <- matrix(sample(1:9, 9, rep=F), nrow=3)

#### Function to minimize

The function to minimize is critically important. A good idea is to create a function that first calculates the 8 row/column/diagonal sums. Then for each of these values, determine (in absolute) value, how far they are from 15. The return value can be the mean (or max) of these distances.

#### Proposal transition

It might seem sufficient to just swap entries around, but it turns out that it’s easy to get stuck in a bad square for which no swap will really improve things. A better idea is to randomly either

• Swap sites: Pick two sites in the $$3\times 3$$ grid and just flip the entries. For example, you could pick the (1,2) site to swap with the (3,3)sites. That means,

$\begin{array}{|c|c|c|} \hline 1&\mathbf{2}&5\\ \hline 3&7&6\\ \hline 8&4&\mathbf{9}\\ \hline \end{array}\rightarrow\begin{array}{|c|c|c|} \hline 1&\mathbf{9}&5\\ \hline 3&7&6\\ \hline 8&4&\mathbf{2}\\ \hline \end{array}$

• Swap rows or columns: Pick two rows (or columns) and flip the rows/columns. For example, swap the 1st and 2nd rows.

$\begin{array}{|c|c|c|} \hline 1&2&5\\ \hline 3&7&6\\ \hline 8&4&9\\ \hline \end{array}\rightarrow\begin{array}{|c|c|c|} \hline 3&7&6\\ \hline 1&2&5\\ \hline 8&4&9\\ \hline \end{array}$

Hence your move function will have to make a few choices:

• Do you swap sites or row/columns?.
• If you are swapping sites, which two sites?
• If you are swapping row/columns, then decide on rows versus columns. The decide with pair or rows (or columns) to swap.

I suggest writing two functions: swapSites and swapRowCols, and blending these into your proposal function.