*Memoization

Even when programming in a functional style, abstractions like arrays and hash tables can be extremely useful. The "Hashtbl" module in the OCaml standard library provides a type for hash tables, as well as standard operations. One important use of hash tables is for memoization, in which a previously computed result is stored in the table and retrieved later. Memoization expands on this intuition to obtain a powerful technique for building efficient algorithms, especially in a functional language.

Example: Fibbonacci

As an 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 code:

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

Unfortunately, this code takes exponential time: $\Theta(\phi^{n})$ , where $\phi$ is the golden ratio, $(1 + \sqrt{5})/2$ .

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 indexing keys are integers, so we can use implement this table using an array:

let fibm(n) =
  let memo: int option array = Array.create (n+1) None in
  let rec f_mem(n) =
    match memo.(n) with
    Some result -> result            (* computed already! *)
      | None ->
      let result = if n<2 then 1 else f_mem(n-1) + f_mem(n-2) in
        memo.(n) <- (Some result);   (* record in table *)
        result
  in
    f_mem(n)

The function f_mem defined inside fibm contains the original recursive algorithm, except before doing that calculation it first checks if the result has already been computed and stored in the table in which case it simply returns the result. How do we analyze the running time of this function? The time spent in a single call to f_mem is $O(1)$ if we exclude the time spent in any recursive calls that it happens to make. Now we look for a way to bound the total number of recursive calls by finding some measure of the progress that is being made.

A good choice of progress measure, not only here but also for many uses of memoization, is the number of nonempty entries in the table (i.e. entries that contain Some n rather than None). Each time f_mem makes the two recursive calls it also increases the number of nonempty entries by one (filling in a formerly empty entry in the table with a new value). Since the table has only n entries, there can thus only be a total of $O(n)$ calls to f_mem, for a total running time of $O(n)$ (because we established above that each call takes $O(1)$ time). This speedup from memoization thus reduces the running time from exponential to linear, a huge change---e.g., for $n=4$ the speedup from memoization is more than a factor of a million!

The key to being able to apply memoziation is that there are common subproblems which are being solved repeatedly. Thus we are able to use some extra storage to save on repeated computation.

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

Example: 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:

module Unmemoized = struct
  type tree = Empty | Node of int * tree * tree

  (* Returns optimum fun for t. *)
  let rec party(t) = max (party_in t) (party_out t)

  (* Returns optimum fun for t assuming the root node of t
   * is included. *)
  and party_in(t) =
    match t with 
        Empty -> 0
      | Node(v,left,right) -> v + party_out(left)
          + party_out(right)

  (* Returns optimum fun for t assuming the root node of t
   * is excluded. *)
  and party_out(t) =
    match t with
    Empty -> 0
      | Node(v,left,right) -> party(left) + party(right)
end

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.

module Memoized = struct
  (* This version memoizes the optimal fun value for each tree node.
     It also remembers the best invite list. Each tree node has the
     name of the employee as a string. *)
  type tree = Empty
              | Node of int * string * tree * tree *
                  ((int*string list) option) ref

  let rec party(t): int * string list =
    match t with
        Empty -> (0, [])
      | Node(v,name,left,right,memo) ->
          (match !memo with
               Some result -> result
             | None ->
                 let (infun, innames) = party_in(t) in
                 let (outfun, outnames) = party_out(t) in
                 let result =
           if infun > outfun then (v + infun, name :: innames)
           else (outfun, outnames)
                 in
                   (memo := Some result); result)

  and party_in(t) =
    match t with
        Empty -> (0, [])
      | Node(v,name,l,r,_) ->
          let (lfun, lnames) = party_out(l)
          and (rfun, rnames) = party_out(r) in
            ((v + lfun + rfun), name :: lnames @ rnames)

  and party_out(t) =
    match t with
    Empty -> (0, [])
      | Node(v,_,l,r,_) ->
          let (lfun, lnames) = party(l)
          and (rfun, rnames) = party(r) in
            ((lfun + rfun), lnames @ rnames)
end

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.

Example: 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:

this may be a
difficult
example

Note that the optimal splitting is:

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 Omega(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:

(* Result of formatting a string. A result (lst, n)
   means a string was formatted into the lines in lst,
   with a total sum-of-cubes cost of n. *)
type breakResult = string list * int

(* Result: format the words in "words" into a list of lines optimally,
 * minimizing the sum of the cubes of differences between the line lengths
 * and "target".
 * Performance: worst-case time is exponential in the number of words.
 *)
let linebreak1 (words: string list)
    (target: int): string list =
  let rec lb(clen: int) (words: string list): breakResult =
    match words with
    [] -> ([""], 0) (* no charge for last line *)
      | word::rest ->
      (* Try two ways of doing it: (1) insert a linebreak right after
           * current word, or (2) continue the current line.  Pick the
       * better one.  *)
      let wlen = String.length(word) in
      let contlen = if clen = 0 then wlen else clen + 1 + wlen in
          let (l1, c1') = lb 0 rest in
          let c1 = c1' + cube(target - contlen) in
            if contlen <= target then
              let (h2::t2, c2) = lb contlen rest
              in
        if c1 < c2 then (word::l1, c1)
        else ((if h2="" then word
               else word^" "^h2)::t2, c2)
            else
              (word::l1, big)
  in
  let (result, cost) = lb 0 words in
    result

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:

(* Same spec as linebreak1.
   Performance: worst-case time is linear in the number of words. *)
let linebreak2(words: string list)
    (target: int): string list =
  let memo: breakResult option array = 
                          Array.create (List.length words+1) None in
  let rec lb_mem(words: string list): breakResult =
    let n = List.length words
    in
      match Array.get memo n with
      Some br -> br
    | None -> let br = lb 0 words in
        Array.set memo n (Some br);
            br
  and lb(clen: int) (words: string list): breakResult =
    match words with
    [] -> ([""], 0) (* no charge for last line *)
      | word::rest ->
          let wlen = String.length(word) in
      let contlen = if clen = 0 then wlen else clen + 1 + wlen in
          let (l1, c1') = lb_mem(rest) in
          let c1 = c1' + cube(target - contlen) in
            if contlen > target then
              (word::l1, big)
        else
              let (h2::t2, c2) = lb contlen rest
              in
        if c1 < c2 then (word::l1, c1)
        else ((if h2="" then word
               else word^" "^h2)::t2, c2)
  in
  let (result, cost) = lb 0 words in
    result

Memoization using higher order functions

We can use the higher order functions to memoize any function. First consider the case of memoizing a non-recursive function f. In that case we simply need to create a hash table that stores the corresponding value for each argument that f is called with (and to memoize multi-argument functions we can use currying and uncurrying to convert to a single argument function).

let memo f =
  let h = Hashtbl.create 11 in
    fun x ->
      try
        Hashtbl.find h x
      with
          Not_found ->
            let y = f x in
              Hashtbl.add h x y;
              y

For recursive functions, however, the recursive call structure needs to be modified. This can be abstracted out independent of the function that is being memoized:

let memo_rec f =
  let h = Hashtbl.create 11 in
  let rec g x =
    try
      Hashtbl.find h x
    with
        Not_found ->
          let y = f g x in
            Hashtbl.add h x y ;
            y
  in
    g

Now we can slightly rewrite the original fib function from the beginning of lecture using this general memoization technique:

let fib_memo =
  let rec fib self = function (n) -> 
    if n<2 then 1 else self(n-1) + self(n-2)
  in
    memo_rec fib

Conclusion

Memoization is a powerful technique for asymptotically speeding up simple recursive algorithms, without having to change the way the algorithm works. Memoization can be seen as an instance of the extremely useful algorithmic technique called dynamic programming, which you will see in CS 4820.

12.5. Memoization