What is Dynamic Programming?
Dynamic programming is an algorithmic optimization technique that breaks down a complicated problem into smaller overlapping sub-problems in a recursive manner and uses solutions to the sub-problems to construct a solution to the original problem.
"Dynamic programming”, what an awfully scary name. What does it even mean?? What’s so “dynamic” about programming?
The name was invented by Richard Bellman in the 1950s when computers were still decades away. So by “programming” he did NOT mean programming as coding at a computer. Bellman was a mathematician, and what he really meant by programming was “planning” and “decision making”.
Trivia time: according to Wikipedia, Bellman was working at RAND corporation, and it was hard to get mathematical research funding at the time. To disguise the fact that he was conducting mathematical research, he phrased his research in a less mathematical term, “dynamic programming”. “Bellman chose the word dynamic to capture the time-varying aspect of the problems and because it sounded impressive. The word programming referred to the use of the method to find an optimal program, in the sense of a military schedule for training or logistics.”
So he really meant “multistage planning”, a simple concept of solving bigger problems using smaller problems while saving results to avoid repeated calculations. That sounds awfully familiar. Isn’t that memoization? Yes, it is. Keep on reading.
Characteristics of Dynamic Programming
A problem is a dynamic programming problem if it satisfies two conditions:
The problem can be divided into sub-problems, and its optimal solution can be constructed from optimal solutions of the sub-problems. In academic terms, this is called optimal substructure.
The sub-problems from 1) overlap.
1. Optimal substructure
Consider the problem of the shortest driving path from San Francisco (SF) to San Diego (SD). Since the highway goes through Los Angeles (LA), the problem can be divided into two sub-problems - driving from SF to LA and driving from LA to SD.
In addition, shortest_path(SF, SD) = shortest_path(SF, LA) + shortest_path(LA, SD). Optimal solution to the problem = combination of optimal solutions of the sub-problems.
Now let’s look at an example where the problem does NOT have an optimal substructure. Consider buying the cheapest airline ticket from New York (NYC) to San Francisco (SF). Let’s assume there is no direct flight, and we must transit through Chicago (CHI). Even though our trip is divided into two parts, NYC to CHI and CHI to SF, usually the cheapest ticket from NYC to SF != the cheapest ticket from NYC to CHI + the cheapest ticket from CHI to SF because airlines do generally not price multi-leg trips the sum of each flights to maximize profit.
2. Overlapping sub-problems
As we have seen in the memoization section, Fibonacci number calculation has a good amount of repeated computation (overlapping sub-problems) whose results can be cached and reused.
If the two conditions stated above are satisfied, then dynamic programming can solve the problem.
DP == DFS + memoization + pruning
You might have seen posts on the coding forum titled “simple DFS solution” and “0.5 sec DP solution” for the same problem. It is because the two methods are equivalent. There are two different approaches to DP—top-down and bottom-up.
It's important to mention that pruning is an integral part of the process to cut down run time. We have seen state-space tree branch pruning in the backtracking section. We will see how it's applied in the following Knapsack section.
How to Solve Dynamic Programming Problems?
Top-down: this is basically DFS + memoization as we have seen memoization. We split large problems and recursively solve smaller sub-problems.
Bottom-up: we try to solve sub-problems and then use their solutions to find the solutions to bigger sub-problems. This is usually done in a tabular form.
Let’s look at a concrete example.
Let's revisit the Fibonacci number problem from the memoization section.
Top-down with Memoization
Recall that we have a system for backtracking and memoization.
Draw the tree: see the tree above
- What state do we need to know if we have reached a solution?
We need to know the value of
nwe are computing.
- What state do we need to decide which child nodes to visit next?
No extra state is required. We always visit
- DFS + pruning (if needed) + memoization
1def fib(n, memo): 2 if n in memo: # check for the solution in the memo, if found, return it right away 3 return memo[n] 4 5 if n == 0 or n == 1: 6 return n 7 8 res = fib(n - 1, memo) + fib(n - 2, memo) 9 10 memo[n] = res # save the solution in memo before returning 11 return res
Bottom-up with Tabulation
For bottom-up dynamic programming, we first want to start with the sub-problems and work our way up to the main problem. This is usually done by filling up a table.
For the Fibonacci problem, we want to fill a one-dimensional table
dp, where each entry at index
i represents the value of the Fibonacci number at index i. The last element of the array is the result we want to return.
The order of filling matters because we cannot calculate
dp[i - 1] and
dp[i - 2].
1def fib(n): 2 dp = [0, 1] 3 for i in range(2, n + 1): 4 dp.append(dp[i - 1] + dp[i - 2]) 5 6 return dp[-1]
sub-problems and Recurrence Relation
dp[i] = dp[i - 1] + dp[i - 2] is called the recurrence relation. It is the key to solving any dynamic programing problem.
For the Fibonacci number problem, the relation is already given
dp[i] = dp[i - 1] + dp[i - 2]. We will discuss the patterns of recurrence relation in the next section.
Should I do top-down or bottom-up?
- The order of computing sub-problems doesn't matter. For bottom-up, we have to fill the table in order to solve all the sub-problems first. For example, to fill
dp, we have to have filled
dpfirst. For top-down, we can let recursion and memoization take care of the sub-problems and, therefore, not worry about the order.
- Easier to the reason for partition type of problems (how many ways are there too.., splitting a string into...). Just do DFS and add memoization.
- Easier to analyze the time complexity (since it's just the time to fill the table)
- No recursion, and thus no system stack overflow—although not a huge concern for normal coding interviews.
From our experiences, top-down is often a better place to start unless it's a grid problem where the states are in plain sight.
Greedy Algorithm vs. Dynamic Programming
What is a greedy algorithm? As the name suggests, it is an algorithm where we always want to choose the best answer. The main difference between a greedy algorithm and dynamic programming is that the answer to a dynamic programming problem is not always necessarily the best answer for every state. This can be due to other restrictions in the problem statement, such that we don't always want to pick the best answer. A good way to distinguish between the two is to figure out a dynamic programming solution and see if you can optimize it by always picking the best answer for the dynamic programming sub-states.
For example, given a series of intervals, you were asked to pick the minimum number of intervals required to cover a given length.
dp[i] denote the minimum number of intervals to make an interval of length
dp[i] = min(dp[i], dp[i - length[j]] + 1) where
length is the array containing the interval lengths. We then realize that for our
dp state that we should greedily pick the longest interval each time if permitted, which leads us to our greedy solution.
This is a rather simple example, but it may be helpful for more obscure greedy solutions disguised as dynamic programming problems.
Divide and Conquer vs. Dynamic Programming
Both Divide and Conquer and dynamic programming break the original problem down into multiple sub-problems. The difference is that in dynamic programming, the sub-problems overlap, whereas in divide and conquer, they don't.
Consider Merge Sort, the sub-arrays are sorted and merged, but the sub-arrays do not overlap. Now consider Fibonacci; the green and red nodes in the "overlapping sub-problems" overlap.
When to use dynamic programming
Mathematically, dynamic programming is an optimization method on one or more sequences (e.g., arrays, matrices). So questions asking about the optimal way to do something on one or more sequences are often a good candidate for dynamic programming. Signs of dynamic programming:
- The problem asks for the maximum/longest, minimal/shortest value/cost/profit you can get from doing operations on a sequence.
- You've tried greedy, but it sometimes gives the wrong solution. This often means you have to consider sub-problems for an optimal solution.
- The problem asks how many ways there are to do something. This can often be solved by DFS + memoization, i.e., top-down dynamic programming.
- Partition a string/array into sub-sequences so that a specific condition is met. This is often well-suited for top-down dynamic programming.
- The problem is about the optimal way to play a game.
How to Develop Intuition for Dynamic Programming Problems
As you may have noticed, the concept of DP is quite simple—find the overlapping sub-problems, solve them, and use the sub-problem solutions to find the solution to the original problem. The hard part is to know how to find the recurrence relation. To best way to develop intuition is to get familiar with common patterns. Some classic examples include longest common subsequence (LCS), 0-1 knapsack, and longest increasing subsequence (LIS).
Dynamic Programming Patterns
Here's the breakdown. We also highlighted the keywords that indicate it's likely a dynamic programming problem.
This is the most common type of DP problem and an excellent place to get a feel of dynamic programming and how it's different from brute force backtracking. The state in these problems is a two-variable pair instead of the single-variable state we have seen so far in backtracking.
- Knapsack - given a number of items of different weights, is it possible to use the items to make up weight X?
- Partition an array into two equal sum subsets - is it possible to divide an array into two subsets with equal sum?
Note that we categorize this section as "weight-only" knapsack to differentiate the classic textbook 0-1 knapsack.
- 0-1 Knapsack - same as weight-only knapsack except items have values, and the goal is to find the maximum object value we can put in our knapsack without exceeding the allowed weight.
The state in this type of DP is often the grid itself.
dp[i][j] means max/min/best value for matrix cell ending at index i, j.
- Robot unique paths - number of ways for robot to move from top left to bottom right
- Min path sum - find path in a grid with minimum cost
- Maximal square - find maximal square of 1s in a grid of 0s and 1s
This type of problem asks whether a player can win a decision game. The key to solving game theory problems is to identify a winning state, and formulate a winning state as a state that returns a losing state to the opponent
- Divisor game These problems are often closely related to the following Interval DP problems.
The key to solving this type of problem involves finding a sub-problem defined on an interval
- Coin game - two players play a game by removing coins from either end of a row of coins. Find the maximum store.
- Festival game, bursting balloons - similar to the coin game problem but with a different way of evaluating scores.
This type of problem has two sequences in its problem statement.
dp[i][j] represents the max/min/best value for the first sequence ending in index i and the second sequence ending in index j.
- Longest common subsequence - find the longest common subsequence that is common in two sequences
- Edit distance - find the minimum distance to edit one string to another
Dynamic number of sub-problems, Longest Increasing Subsequence
This type of DP problem is unique in that the current state depends on a dynamic number of previous states, e.g.
dp[i] = max(d[j]..) for j from 0 to i.
- Longest Increasing Subsequence - find the longest increasing subsequence of an array of numbers
- Buy/sell the stock with at most K transactions - maximize profit by buying and selling stocks using at most K transaction
These DP problems use bitmasks to reduce factorial complexity (n!) to 2^n by encoding the dp state in bitmasks.
Still not clear? Submit the part you don't understand to our editors.