- If the current colour of a square is white then if it is surrounded by 2 or 3 adjacent white squares it stays white, otherwise it becomes black.
- If the current colour of a square is black then if it is surrounded by exactly 3 white squares then it becomes white, otherwise it stays black.
If we treat the grid as a numeric matrix where 0 is black and 1 is white, then a differentiable version of these rules can e made. The idea here is to allow not just black and white but also shades of grey so that the values of the squares can change only slightly and hence allow the changes to be differentiable. In order to force the values to remain between 1 and 0 we will make use of the sigmoid (logistic) function which is defined as $y = \frac{1}{1 + e^{-x}}$.
The key trick here is to come up with a differentiable function that maps the number of white squares around a square to 0 or 1. We need two such functions: one that when x is 2 or 3 y is 1 and otherwise y is 0 and another that when x is 3 y is 1 and otherwise y is 0. We can make this function using two sigmoid functions subtracted from each other as follows:
$y = \text{sig}(w \times (x - a)) - \text{sig}(w \times (x - b))$
The graph plot of this function would be one which starts as 0, then increases to 1 at $x = a$ (where $a$ is the halfway point as $y$ goes from 0 to 1), then stays 1 until $x = b$, at which point $y$ goes back to 0 (again, $b$ is the halfway point as $y$ goes from 1 to 0). $w$ is there to control how steep the change from 0 to 1 and back should be with large $w$ meaning very steep.
So now we apply the above equation to be used on a whole matrix of values and where $x$ is the number of adjacent white squares. Given that a square can also be grey, we instead just take the sum of the adjacent values to count the number of white squares. This results in a count which is not a whole number but which is usable in a differentiable function.
We also need to be able to choose between using the rule for white squares or the rule for black squares. To do that we can just take a weighted average as follows:
$y = x \times a + (1 - x) \times b$
where $x$ is a number between 0 and 1 and a is the value $y$ should be when $x$ is 1 whilst $b$ is the value $y$ should be when $x$ is 0.
Given a matrix $G$, you can get the next iteration $G'$ using the following function:
Let $k$ be a 3 by 3 matrix consisting of all 1s and a single 0 in the center.
$N = conv(G, k)$, that is, convolve the kernel $k$ over the matrix $G$ in order to get a new matrix $N$ giving the number of adjacent white squares around every element in the matrix.
$B = \text{sig}(w \times (G - 2.5)) - \text{sig}(w \times (G - 3.5))$ (resultant matrix if all elements where black)
$W = \text{sig}(w \times (G - 1.5)) - \text{sig}(w \times (G - 3.5))$ (resultant matrix if all elements where white)
$G' = G \times W + (1 - G) \times B$
