Lecture 25: Dynamic Programming: Matlab Code

University of Southern California

Linguistics 285

USC Linguistics

December 1, 2015

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 1/1
Dynamic Programming Approach

I Dynamic Programming is an alternative search strategy that is faster

than Exhaustive search, slower than Greedy search, but gives the
optimal solution.
I View a problem as consisting of subproblems:
I Aim: Solve main problem
I To achieve that aim, you need to solve some subproblems
I To achieve the solution to these subproblems, you need to solve a set
of subsubproblems
I And so on...
I Dynamic Programming works when the subproblems have similar
forms, and when the tiniest subproblems have very easy solutions.

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 2/1
Dynamic Programming

B D G

A
E H J
C
F I

I So we now have 2 subproblems VBJ and

I Main Problem: Shortest path VCJ . If we could solve those subproblems,
from A to J: VAJ we could solve the main problem VAJ .
I DP Thinking: Let’s say I know I But now we can think of VBJ as its own
the best paths from B to J and problem, and then repeat the thinking: If I
C to J: VBJ and VCJ . Then we knew the solutions to VDJ , VEJ , and VFJ ,
would add VBJ to the cost from then we can solve VBJ . Same with VCJ .
A to B, VCJ to the cost from A I Continue thinking in this way till we get to:
to C, compare the two, and VGJ , VHJ , VIJ , which are easy to solve!
pick the least.
Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 3/1
Dynamic Programming

B 5
D 8
G
9
12
6 7 5
A 14
7
9
E 10 10
H 9
J
7
10
C 9
7 8
11
F 8
I
8 9
< 8 + VGJ =
8 9 VDJ = min 7 + VHJ
: ;
<5 + VDJ = 10 + VIJ
VBJ = min 6 + VEJ VGJ = 5
: ; 8 9
⇢ 9 + VF J < 9 + VGJ =
12 + VBJ VHJ = 9
VAJ = min VEJ = min 7 + VHJ
7 + VCJ 8 9 : ;
<14 + VDJ = 9+V IJ
VCJ = min 10 + VEJ VIJ = 8
: ; 8 9
11 + VF J <10 + VGJ =
VF J = min 7 + VHJ
: ;
8+V IJ

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 4/1
Dynamic Programming solution in Matlab
2, 1 5
3, 1 8
4, 1
9
12
6 7 5
1, 1 14
7
9
3, 2 10 10
4, 2 9
5, 1
7
10
2, 2 9
7 8
11
3, 3 8
4, 3
8 9
< 8 + VGJ =
8 9 VDJ = min 7 + VHJ
: ;
<5 + VDJ = 10 + VIJ
VBJ = min 6 + VEJ VGJ = 5
: ; 8 9
⇢ 9+V FJ < 9 + VGJ =
12 + VBJ VHJ = 9
VAJ = min VEJ = min 7 + VHJ
7 + VCJ 8 9 : ;
<14 + VDJ = 9 + VIJ
VCJ = min 10 + VEJ VIJ = 8
: ; 8 9
11 + VF J <10 + VGJ =
VF J = min 7 + VHJ
: ;
8 + VIJ

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 5/1
Dynamic Programming notation: distances

I cost (or distance) of going from stage 1, state 1 to stage 2, state 1

I d(1, 1, 2, 1)
I cost (or distance) of going from stage 1, state 1 to stage 2, state 2
I d(1, 1, 2, 2)
I cost (or distance) of going from stage 2, state 1 to stage 3, state 2
I d(2, 1, 3, 2)
I cost (or distance) of going from stage k, state i to stage k+1, state j
I d(k, i, k + 1, j)

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 6/1
Dynamic Programming notation: Minimum cost from a
node to the end

I minimum cost of going from stage 4, state 1 to end

I V (4, 1)
I minimum cost of going from stage 3, state 2 to end
I V (3, 2)
I Evaluating:
I V (4, 1) = 5
 
d(3, 2, 4, 1) + V (4, 1)
I V (3, 2) = min d(3, 2, 4, 2) + V (4, 2)
 
d(3, 2, 4, 3) + V (4, 3)
I Generalizing:
I V (k, i) = minj (d(k, i, k + 1, j) + V (k + 1, j))

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 7/1
Matlab code for DP: Defining the network

costs(1,1,2,1) = 12;
costs(1,1,2,2) = 7;
2, 1 5
3, 1 8
4, 1 costs(2,1,3,1) = 5;
costs(2,1,3,2) = 6;
costs(2,1,3,3) = 9;
9 costs(2,2,3,1) = 14;
12 costs(2,2,3,2) = 10;
6 7 5 costs(2,2,3,3) = 11;
1, 1 14
costs(3,1,4,1) = 8;
7 costs(3,1,4,2) = 7;

9
3, 2 10 10
4, 2 9
5, 1 costs(3,1,4,3) = 10;
7 costs(3,2,4,1) = 9;
costs(3,2,4,2) = 7;
10
2, 2 9 costs(3,2,4,3) = 9;
costs(3,3,4,1) = 10;
7 8 costs(3,3,4,2) = 7;
11 costs(3,3,4,3) = 8;
costs(4,1,5,1) = 5;
3, 3 8
4, 3 costs(4,2,5,1) = 9;
costs(4,3,5,1) = 8;
num_states = [1 2 3 3 1];

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 8/1
Matlab code for DP
I Using this generalized form, we can write a Matlab program, using
nested loops, that will start at the end and compute V (k, i) for every
node recursively.
I The last one we compute will be V (1, 1) which is the length of the
minimum path from beginning to end.
d = costs;
V(5,1)=0;
for k=4:-1:1
for i=1:num_states(k)
for j = 1:num_states(k+1)
path_length(j)= d(k,i,k+1,j)+V(k+1,j);
end
V(k,i)=min(path_length);
clear path_length
end
end
V(1,1)

Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 9/1
Matlab code for DP
I Why is this incomplete?
I We know the minimum length path, but we don’t know which states
it passes through.
I Now start at the begining. Add the cost of going from stage k to
each of the nodes at stage k + 1.
I Find which total is minimal and choose the corresponding state in
stage k + 1.
path = 0;
index = 1;
for k=1:4
for j=1:num_states(k+1)
path_length(j)= d(k,index,k+1,j)+V(k+1,j);
end
[minval, index] = min(path_length);
path(k) = index;
clear path_length
end
path
Linguistics 285 (USC Linguistics) Lecture 25: Dynamic Programming: Matlab Code December 1, 2015 10 / 1