Introduction

Reading:
Dynamic Programming I “Dynamic Programming”, 15 CLRS
I “Dynamic Programming”, 5.3 GT
Jonathan Backer
Greedy works if some optimal solution contains the greedy choice.
[email protected]
I Dijkstra’s algorithm always adds the cheapest vertex to the
shortest path tree (greedy).
Department of Computer Science I Dijkstra’s algorithm may not work with negative edge weights.
University of British Columbia

Dynamic programming tries all possible choices.

I Bellman-Ford’s algorithm attempts every one edge extension
July 5, 2007
of shortest paths (exhaustive).
I Bellman-Ford’s algorithm works with negative edge weights.

Optimal substructure Making change with coins

Both Dijkstra’s and Bellman-Ford’s algorithms work because you

Problem
can extend the optimal solution of a subproblem.
Given: Coin values c1 , c2 , . . . , ct with which to make change and
Definition the amount of change to be made n.
A problem has optimal substructure if some optimal solution is Wanted: Number of each coin to use n1 , n2 , . . . , nt such that sum
I an optimal solution to a subproblem combined with of coins is n and fewest coins are used.
I an optimal choice. Denominations chosen so that greedy algorithm works, but not
true in general.
I Often don’t know which choice to make, so try them all.
Example
I May be efficient if subproblems overlap
Coins: 1¢, 3¢, and 4¢ Greedy → 4¢, 1¢, and 1¢
I |V | paths to extend in Bellman-Ford (subproblems)
I |E | edges to extend with (choices) Change to make: 6¢ Optimal → 3¢ and 3¢
Exhaustive coin changing Recursion tree for TryEmAll

Algorithm TryEmAll(C ,n) This is inefficient because it recomputes the same subproblems
int N[C .length] over and over again.
for i ← 0 to N.length−1 do
N[i] ← 0 TryEmAll([1,3,4],90)
90
if n = 0 then
return N
89 87 86
N[1] ← ∞
for i ← 0 to C .length−1 do
88 86 85 86 84 83 85 83 82
if n ≥ C [i] then
subprob ← TryEmAll(C ,n − C [i])
A better idea: replace each recursive call with a table look-up.
if subprob.sum()+1 < N.sum() then
N ← subprob I Construct a table to store the optimal solution for each n.
N[i] ← N[i] + 1 I Iteratively increase n and compute its entry.
return N

Dynamic programming solution Runtime complexity

Algorithm DPCoinChange(C ,n) What is the runtime complexity of this algorithm?

int N[n + 1][C .length]
I If it updates N[m] every time, then n × t updates.
I Each update copies t integers.
for i ← 0 to C .length−1 do
N[0][i] ← 0
I So O(nt 2 ).

for m ← 1 to n do Advantages of eliminating the recursion:

N[m][0] ← ∞
I Counting argument for runtime complexity.
for i ← 0 to C .length−1 do
if m ≥ C [i] then I No call stack overhead!
if N[m − C [i]].sum()+1 < N[m].sum() then
N[m] ← N[m − C [i]] Why copy solution during update when we only chose one coin?
N[m][i] ← N[m][i] + 1 I Faster to remember the optimal choice and
return N[n] I backtrack to recover the solution.
Faster solution Backtracking
// backtracking
int N[C .length]
Algorithm FastDPCoinChange(C ,n) for i ← 0 to N.length−1 do
int minCoins[n + 1], bestChoice[n + 1] N[i] ← 0
minCoins[0] ← 0 while n > 0 do
bestChoice[0] ← −1 N[bestChoice[n]] ← N[bestChoice[n]] + 1
// main loop n ← n − C [bestChoice[n]]
for m ← 1 to n do return N
minCoins[0] ← ∞
for i ← 0 to C .length−1 do Eliminates copying and adding t integers from the innermost loop.
if m ≥ C [i] then I Total runtime complexity is O(nt).
if minCoins[m − C [i]] + 1 < minCoins[m] then
minCoins[m] ← minCoins[m − C [i]] + 1 Example: 1¢, 3¢, and 4¢ coins
bestChoice[m] ← i n 0 1 2 3 4 5 6 7 8 9 10 How do you
N 0 1 2 1 1 2 2 2 2 make 9¢ and
bC ∅ 1 1 3 4 1 3 3 4 10¢ change?

Designing a dynamic programming algorithm Design (cont’d)

Decide on the parameters the problem will have.

I This gives the “shape” of the table and determines the Express the problem in terms of smaller problems.
runtime complexity. I FastDPCoinChange
I FastDPCoinChange only used n to determine bestChoice, so
table is an array. minCoins[n] = min{1 + minCoins[n − C [i]] : C [i] ≤ n}
I Limited supply of coins
I Solution depends on n and number of each type of coin Determine how to fill-in the table.
available (a1 , a2 , . . . , at ). I A subproblem solution must be computed before those that
I Table has one dimension for n, another for a1 , another for a2 , rely on it.
etc.
I Trickier for multi-dimensional tables. Typically row-by-row,
What do we need to store in the table: column-by-column, or diagonal-by-diagonal.
I DPCoinChange stored all of the best choices made so far.
I FastDPCoinChange just stored the last best choice.
Memoization (top-down)

Use divide-and-conquer to fill-in the table.

I Return value if already computed.
I Recurse otherwise.
I Save solution in table before returning.
Pros:
I If some subproblem is irrelevant, memoization won’t solve it.
I If you cannot figure out how to fill the table,
divide-and-conquer will do it for you.
Cons:
I Recursive function calling overhead (stack frame).
I Sometimes miss tricks like we used in FastDPCoinChange.