Dynamic programming is a fancy name for using divide-and-conquer technique with a table. As compared to divide-and-conquer, dynamic programming is more powerful and subtle design technique. Let me repeat , it is not a specific algorithm, but it is a meta-technique (like divide-and-conquer). This technique was developed back in the days when "programming" meant "tabular method" (like linear programming). It does not really refer to computer programming. Here in our advanced algorithm course, we'll also think of "programming" as a "tableau method" and certainly not writing code. Dynamic programming is a stage-wise search method suitable for optimization problems whose solutions may be viewed as the result of a sequence of decisions. The most attractive property of this strategy is that during the search for a solution it avoids full enumeration by pruning early partial decision solutions that cannot possibly lead to optimal solution. In many practical situations, this strategy hits the optimal solution in a polynomial number of decision steps. However, in the worst case, such a strategy may end up performing full enumeration.

Dynamic programming takes advantage of the duplication and arrange to solve each subproblem only once, saving the solution (in table or in a globally accessible place) for later use. The underlying idea of dynamic programming  is: avoid calculating the same stuff twice, usually by keeping a table of known results of subproblems. Unlike divide-and-conquer, which solves the subproblems top-down, a dynamic programming is a bottom-up technique. The dynamic programming technique is related to divide-and-conquer, in the sense that it breaks problem down into smaller problems and it solves recursively. However, because of the somewhat different nature of dynamic programming problems, standard divide-and-conquer solutions are not usually efficient.

The dynamic programming is among the most powerful for designing algorithms for optimization problem. This is true for two reasons. Firstly, dynamic programming solutions are based on few common elements. Secondly, dynamic programming problems are typical optimization problems i.e., find the minimum or maximum cost solution, subject to various constraints.

In other words, this technique used for optimization problems:

• Find a solution to the problem with the optimal value.
• Then perform minimization or maximization. (We'll see example of both in CLRS).

The dynamic programming is a paradigm of algorithm design in which an optimization problem is solved by a combination of caching subproblem solutions and appealing to the "principle of optimality."

There are three basic elements that characterize a dynamic programming algorithm:

1. Substructure

Decompose the given problem into smaller (and hopefully simpler) subproblems. Express the solution of the original problem in terms of solutions for smaller problems. Note that unlike divide-and-conquer problems, it is not usually sufficient to consider one decomposition, but many different ones.

2. Table-Structure

After solving the subproblems, store the answers (results) to the subproblems in a table. This is done because (typically) subproblem solutions are reused many times, and we do not want to repeatedly solve the same problem over and over again.

3. Bottom-up Computation

Using table (or something), combine solutions of smaller subproblems to solve larger subproblems, and eventually arrive at a solution to the complete problem. The idea of bottom-up computation is as follow:

Bottom-up means

1. Start with the smallest subproblems.
2. Combining theirs solutions obtain the solutions to subproblems of increasing size.
3. Until arrive at the solution of the original problem.

Once we decided that we are going to attack the given problem with dynamic programming technique, the most important step is the formulation of the problem. In other words, the most important question in designing a dynamic programming solution to a problem is how to set up the subproblem structure.

If I can't apply dynamic programming to all optimization problem, then the question is what should I look for to apply this technique? Well! the answer is there are two important elements that a problem must have in order for dynamic programming technique to be applicable (look for those!).

1. Optimal Substructure  

Show that a solution to a problem consists of making a choice, which leaves one or sub-problems to solve. Now suppose that you are given this last choice to an optimal solution. [Students often have trouble understanding the relationship between optimal substructure and determining which choice is made in an optimal solution. One way to understand optimal substructure is to imagine that "God" tells you what was the last choice made in an optimal solution.] Given this choice, determine which subproblems arise and how to characterize the resulting space of subproblems. Show that the solutions to the subproblems used within the optimal solution must themselves be optimal (optimality principle). You usually use cut-and-paste:

• Suppose that one of the subproblem is not optimal.
• Cut it out.
• Paste in an optimal solution.
• Get a better solution to the original problem. Contradicts optimality of problem solution.

That was optimal substructure.

You need to ensure that you consider a wide enough range of choices and subproblems that you get them all . ["God" is too busy to tell you what that last choice really was.] Try all the choices, solve all the subproblems resulting from each choice, and pick the choice whose solution, along the subproblem solutions, is best.

We have used "Optimality Principle" a couple of times. Now a word about this beast: The optimal solution to the problem contains within it optimal solutions to subproblems. This is some times called the principle of optimality.

The Principle of Optimality

The dynamic programming  relies on a principle of optimality. This principle states that in an optimal sequence of decisions or choices, each subsequence must also be optimal. For example, in matrix chain multiplication problem, not only the value we are interested in is optimal but all the other entries in the table are also represent optimal. The principle can be related as follows: the optimal solution to a problem is a combination of optimal solutions to some of its subproblems. The difficulty in turning the principle of optimally into an algorithm is that it is not usually obvious which subproblems are relevant to the problem under consideration.

Now the question is how to characterize the space of subproblems?

• Keep the space as simple as possible.
• Expand it as necessary.

As an example, consider the assembly-line scheduling. In this problem, space of subproblems was fastest way from factory entry through stations S_{1, j}  and S_{2, j}. Clearly, no need to try a more general space of subproblems. On the hand, in case of optimal binary search trees. Suppose we had tried to constrain space of subproblems to subtrees with keys k₁, k₂, . . . , k_j. An optimal BST would have root k_r , for some 1 ≤ r ≤ j. Get subproblems k₁, . . . , k_{r − 1} and k_{r + 1}, . . . , k_j. Unless we could guarantee that r = j, so that subproblem with k_{r + 1}, . . . , k_j is empty, then this subproblem is not of the form k₁, k₂, . . . , k_j. Thus, needed to allow the subproblems to vary at both ends, i.e., allow both i and j to vary.

Optimal substructure varies across problem domains:

1. How many subproblems are used in an optimal solution.
2. How many choices in determining which subproblem(s) to use.

In Assembly-line Scheduling Problem: we have 1 subproblem and 2 choices (for S_{i, j} use either S_{1, j − 1} or S_{2, j − 1}). In the Longest Common Subsequence Problem: we have 1 subproblem but as far as choices are concern, we have either 1 choice (if x_i = y_j , LCS of X_{i − 1} and Y_{j − 1}), or 2 choices (if x_i = y_j , LCS of X_{i − 1} and Y , and LCS of X and Y_{j − 1}). Finally, in case of the Optimal Binary Search Tree Problem: we have 2 subproblems (k_i , . . . , k_{r − 1} and k_{r + 1}, . . . , k_j ) and j − i + 1 choices for k_r in k_i, . . . , k_j . Once we determine optimal solutions to subproblems, we choose from among the j − i + 1 candidates for k_r .

Informally, the running time of the dynamic programming algorithm depends on the overall number of subproblems times the number of choices. For example, in the assembly-line scheduling problem, there are Θ(n) subproblems and 2 choices for each implying running time is Θ(n). In case of longest common subsequence problem, there are Θ(mn) subproblems and at least 2 choices for each implying Θ(mn) running time. Finally, in case of optimal binary search tree problem, we have Θ(n²) sub-problems and Θ(n) choices for each implying Θ(n³) running time.

Dynamic programming uses optimal substructure bottom up fashion:

• First find optimal solutions to subproblems.
• Then choose which to use in optimal solution to the problem.

When we look at greedy algorithms, we'll see that they work in top down fashion:

• First make a choice that looks best.
• Then solve the resulting subproblem.

Warning! You'll surely make an ass out of yourself into thinking optimal substructure applies to all optimization problems. IT DOES NOT. Let me repeat, dynamic programming is not applicable to all optimization problems.
To see this point clearly, go through pages 341 − 344 of CLRS where authors discussed two problems that look similar: Shortest Path Problem and Longest Simple Path Problem. In both problems, they gave us an unweighted, directed graph G = (V, E). And our job is to find a path (sequence of connected edges) from vertex u in V to vertex v in V.

Subproblems Dependencies

It is easy to see that the subproblems, in our above examples, are independent subproblems: For example, in the assembly line problem, there is only 1 subproblem so it is trivially independent. Similarly, in the longest common subsequence problem, again we have only 1 subproblem thus it is automatically independent. On the other hand, in the optimal binary search tree problem, we have two subproblems, k_i, . . . , k_{r − 1} and k_{r + 1}, . . . , k_j, which are clearly independent.

2. Polynomially many (Overlapping) Subproblems

An important aspect to the efficiency of dynamic programming is that the total number of distinct sub-problems to be solved should be at most a polynomial number. Overlapping subproblems occur when recursive algorithm revisits the same problem over and over. A good divide-and-conquer algorithm, for example the merge-sort algorithm, usually generate a brand new problem at each stage of recursion. Our Textbook CLRS has a good example for matrix-chain multiplication to depict this idea. The CLRS also talked about the alternative approach so-called memoization. It works as follows:

• Store, don't recompute
• Make a table indexed by subproblem.
• When solving a subproblem:
  • Lookup in the table.
  • If answer is there, use it.
  • Otherwise, compute answer, then store it.

In dynamic programming, we go one step further. We determine in what order we would want to access the table, and fill it in that way.

Four-Step Method of CLRS

Our Text suggested that the development of a dynamic programming algorithm can be broken into a sequence of following four steps.

1. Characterize the structure of an optimal solution.
2. Recursively defined the value of an optimal solution.
3. Compute the value of an optimal solution in a bottom-up fashion.
4. Construct an optimal solution from computed information.

Updated: March 18, 2010.