Fun with recursion and scoping

Part of the core algorithm in openelec is a routine that finds the best (read: cheapest) way to connect villages. This is based on a minimum spanning tree but, as the MST doesn’t know about the existing grid (or the village populations, or whether there’s a mountain between villages), I had to layer something on top of that. This ‘something’ (discussed here about halfway down the page) starts at currently connected villages and explores outwards along the tree, finding the best configuration that brings electricity to the most people (or the highest demand) with the shortest amount of new wire. Naturally, this calls for a recursive function, which led to a fun exploration of variable scoping.

I find recursion very interesting, as it presents a completely different way of iterating through a task (compared to a for loop) that’s a bit harder to wrap your head around, but often much more intuitive, and sometimes the only way of doing something. As a basic example, consider the function below:

def recurse(keep=0, throw=0):
    keep += 1
    throw += 1
    if keep < 3:
        keep, _ = recurse(keep, throw)
    return keep, throw

recurse()

Which outputs (3, 1). This is because, although both variables have been iterated each time the function is called, throw is not ‘saved’ as control bubbles back up the stack. For a more interesting example, consider the tree below. Imagine that we start at a and want to find the route to (e): a-c-e!

        a
       / \
      b   c
     /   / \
    d  (e)  f

Let’s start by creating the objects to represent our tree. Each node is its own object, and all it knows is whether it’s the target, and who its children are. And we give them names to make our lives easier.

class Node:
    def __init__(self, name, target=False):
        self.children = []
        self.name = name
        self.target = target
    def add(self, *new):
        self.children.extend(new)

a = Node('a')
b = Node('b')
c = Node('c')
d = Node('d')
e = Node('e', True)
f = Node('f')
a.add(b, c)
b.add(d)
c.add(e, f)

Then we’re ready to use a recursive function to find the solution. The first attempt is shown below. Every time the function is called, it adds the current node to current and then explores downstream from there. If the target is found, it copies current into best. Thus we expect that best will only contain the nodes that lead to the target, and the rest will never have bubbled back up the stack (as current is never returned).

def explore(node, current=[], best=[]):
    current.append(node.name)
    if node.target:
        best = current
    for child in node.children:
        best = explore(child, current, best)
    return best

explore(a)

But this outputs ['a', 'b', 'd', 'c', 'e', 'f'] (that is, every node). This is because, in Python, assignment with = doesn’t create objects; it creates a binding between a name and an object. With immutable objects (2, for example, will always be 2), this does what we expect. But with mutable objects (such as lists), even though the scope would suggest we’re dealing with a new variable, in actual fact they’re all pointing to the same underlying object. As a demonstration, consider the following:

foo = [1]
bar = foo
bar.append(2)
print(foo)

This outputs [1, 2], because foo and bar are bound to the same underlying object. We can fix this by replacing bar = foo with bar = foo.copy(), and then we’ll get the output that we expect ([1]). copy gets around this by copying the actual values (and using deepcopy you can go even further). So applying the same logic to our tree-search function, if we add current = current.copy() at the beginning of the function, we can force a locally scoped version of current.

def explore(node, current=[], best=[]):
    current = current.copy()
    current.append(node.name)
    if node.target:
        best = current
    for child in node.children:
        best = explore(child, current, best)
    return best

explore(a)

And voilà, it produces ['a', 'c', 'e'], as expected.

# Addendum

This is still not ideal though: assigning [] as a default parameter is a bad idea, as that object will be created at function definition time and carried around like a confusing nightmare. This is especially true with the list.append() method is used.

The version below fixes these issues. No objects are created at definition time, and everything is treated as immutable (note that in this example, the value of current is never modified).

def explore(node, current=None, best=None):
    if not current:
        current = []
    new_current = current + [node.name]
    if node.target:
        best = new_current
    for child in node.children:
        best = explore(child, new_current, best)
    return best

explore(a)