CS 3110 Lecture 23
Memoization

Even when programming in a functional style, O(1) mutable map abstractions like arrays and hash tables can be extremely useful. One important use of hash tables is for memoization, in which a previously computed result is stored in the table and retrieved later. Memoization is a powerful technique for building efficient algorithms, especially in a functional language.

For example, consider the problem of computing the nth Fibonacci number, defined as f(n) = f(n−1) + f(n−2). We can translate this directly into an SML algorithm:

let f(n) = if n<2 then 1 else f(n-1) + f(n-2)

Unfortunately, this code takes exponential time: Θ(φn), where φ is the golden ratio, (1 + √5)/2. We can easily verify these asymptotic bounds by using the substitution method. Using the recurrence T(n) = T(n−1) + T(n−2) + 1, we show by induction that T(n)≤φn−1:

T(n) = T(n−1) + T(n−2) + 1
kφn−1 − 1 + kφn−2 −1 + 1
kφn−1 + kφn−2 − 1

But φ has the property that φ2 = φ + 1, so:

kφn−1 + kφn−2
= kφn−2 (1 + φ)
= kφn−2 φ2 = kφn

Therefore T(n) ≤ kφn − 1 and T is On). The Ω direction is shown similarly.

The key observation is that the recursive implementation is inefficient because it recomputes the same Fibonacci numbers over and over again. If we record Fibonacci numbers as they are computed, we can avoid this redundant work. The idea is that whenever we compute f(n), we store it in a table indexed by n. In this case the keys are integers, so we can use implement this table using an array:



The function f_alg contains the original recursive algorithm, except that it calls itself via the function f_mem, which checks the table first. In a call to f(n), the function f_alg can be called at most n times, and each of those calls can call f_mem at most twice. So total time is asymptotically linear in n. The speedup from memoization is more than a million at n=40.

Although this code uses imperative constructs (specifically, Array.update), the side effects are not visible outside the function f. Therefore these are benign side effects that do not need to be mentioned in the specification of f.

Party optimization

Suppose we want to throw a party for a company whose org chart is a binary tree. Each employee has an associated “fun value” and we want the set of invited employees to have a maximum total fun value. However, no employee is fun if his superior is invited, so we never invite two employees who are connected in the org chart. (The less fun name for this problem is the maximum weight independent set in a tree.) There are 2n possible invitation lists, so the naive algorithm that compares the fun of every invitation list takes exponential time.

We can use memoization to turn this into a linear-time algorithm. We start by defining a variant type to represent the employees. The int at each node is the fun.

type tree = Empty | Node of int * tree * tree

Now, how can we solve this recursively? One important observation is that in any tree, the optimal invitation list that doesn't include the root node will be the union of optimal invitation lists for the left and right subtrees. And the optimal invitation list that does include the root node will be the union of optimal invitation lists for the left and right children that do not include their respective root nodes. So it seems useful to have functions that optimize the invite lists for the case where the root node is required to be invited, and for the case where the root node is excluded. We'll call these two functions party_in and party_out. Then the result of party is just the maximum of these two functions:



This code has exponential performance. But notice that there are only n possible distinct calls to party. If we change the code to memoize the results of these calls, the performance will be linear in n. Here is a version that memoizes the result of party and also computes the actual invitation lists. Notice that this code memoizes results directly in the tree.



Why was memoization so effective for solving this problem? As with the Fibonacci algorithm, we had the overlapping subproblems property, in which the naive recursive implementation called the function party many times with the same arguments. Memoization saves all those calls. Further, the party optimization problem has the property of optimal substructure, meaning that the optimal answer to a problem is computed from optimal answers to subproblems. Not all optimization problems have this property. The key to using memoization effectively for optimization problems is to figure out how to write a recursive function that implements the algorithm and has two properties. Sometimes this requires thinking carefully.

Optimal line breaking

Here is a more involved example. Suppose that we have some text that we want to format as a paragraph within a certain column width. For example, we might have to do this if we were writing a web browser. For simplicity we will assume that all characters have the same width. A formatting of the text consists of choosing certain pairs of words to put line breaks in between. For example, when applied to the list of words in this paragraph, with width 60, we want output like the following:

let it =
  ["Here is a more involved example of memoization. Suppose that",
   "we have some text that we want to format as a paragraph",
   ...
   "applied to the list of words in this paragraph, with width",
   "60, we want output like the following:"] : string list

A good formatting uses up a lot of each column, and also gives each line similar widths. The greedy approach would be to just fill each line as much as possible, but this can result in lines with very different lengths. For example, if we format the string “this may be a difficult example” at a width of 13 characters, we get a formatting that could be improved:

Greedy Optimal
this may be a
difficult
example
this may be
a difficult
example

The TeX formatting program does a good job of keeping line widths similar by finding the formatting that minimizes the sum of the cube of the leftover space in each line (except for the last). However, for n words, there are Ω(2n) possible formattings, so the algorithm can't possibly check them all for large text inputs. Remarkably, we can use memoization to find the optimal formatting efficiently. In fact, memoization is useful for many optimization problems.

We start by writing a simple recursive algorithm to walk down the list and try either inserting a line break after each word, or not inserting a linebreak:



This algorithm is exponential because it computes all possible formattings. It is therefore much too slow to be practical.

The key observation is that in the optimal formatting of a paragraph of text, the formatting of the text past any given point is the optimal formatting of just that text, given that its first character starts at the column position where the prior formatted text ends. Thus, the formatting problem has optimal substructure when cast in this way.

So if we compute the best formatting after a particular line break position, that formatting is the best for all possible formattings of the text before the break.

We can make linebreak take linear time by memoizing the best formatting for the calls where clen = 0. (We could memoize all calls, but that wouldn't improve speed much. This requires just introducing a function lb_mem that looks up and records memoized formatting results:



Conclusions

Memoization is a powerful technique for asymptotically speeding up simple recursive algorithms, without having to change the way the algorithm works. In general, memoized functions have several arguments, and so hash tables are needed to store the memoized results. Memoization is closely related to the technique of dynamic programming, which you will see in CS482. Dynamic programming requires planning the order in which results are computed, whereas memoization automatically computes results as needed.