Posted by: atri | December 2, 2009

Lect 35: Dyanmic Programming

(Guest post by Shivam Kumar)

The class started with the recap of the various techniques we have learnt so far to design algorithms. Like greedy strategy reduces the exponential runtime to polynomial runtime, and divide and conquer strategy reduces larger polynomial runtime to smaller polynomial runtime ( ex. O(n^2) to O(n\log n).

Then the professor started talking about a new algorithm design technique called Dynamic Programming. Dynamic programming is similar to Divide and Conquer since both use recursion but Dynamic programming uses different kind of recursion. In other words, dynamic programming is smarter about solving recursive sub problems.

End of Semester Blues

If we use greedy strategy on this then we choose jobs (1), (3), and (4) and the total value will be 10 [whereas the schedule that maximizes the value is picking (5) to get a value of 30. –Atri]

Formal Definition of the problem:

  • Input: n jobs, ith job (s_i, f_i, v_i) which stand for start time, finish time, and value respectively.
  • Output: A schedule S\subseteq \{1,\dots,n\} where for i\neq j \in S, i and j are not conflicting.
  • Goal: max \sum_{i\in S} v_i.

If v_i = 1 forall i, then it’s the interval scheduling problem.

Assume: All jobs are sorted by the finish time, i.e., f_1\le f_2 \le \cdots \le f_n. This will be sorted in O(n\log n) time.

Previous non-conflicting jobs” ( p(j) )

  • p(j) = largest i such that i does not conflict with j. (I.e. f_i \le s_j)
  • p(j) = 0 if no such i exists

Ex: compute p values in linear time.

Optimal Schedule for \{1,\dots,j\} will be denoted by \mathcal{O}_j and its value by OPT(j).

Properties of \mathcal{O}_n:

Case 1: n \in \mathcal{O}_n

Question: What can you say about jobs p(n)+1,\dots,n-1 wrt \mathcal{O}_n ?
Answer: None of them will be in \mathcal{O}_n by the definition of p(n).

Claim 1: If  n\in \mathcal{O}_n, then \mathcal{O}_n = (\mathcal{O}_{p(n)}, n).

If p(n) does not conflict with n then p(n-1) will also not conflict.
Recall : jobs are sorted by finish time. This implies that for all i \le p(n), i does not conflict with n. This is because f_i\le f_{p(n)} \le s_n.

Proof of Claim 1: If not, let \mathcal{O}_n = (\mathcal{O}', n) such that  value of  $\mathcal {O}’ >$ value of   \mathcal{O}_{p(n)}. Let \mathcal{O}''=(\mathcal{O}_{p(n)},n). Now, note that

  1. \mathcal{O}''  is a valid schedule
  2. Value of  \mathcal{O}'' <  value  of \mathcal{O}_n.

Contradiction the fact that \mathcal{O}_n is optimal.

Case 2:
n does not belong to \mathcal{O}_n.

Claim 2: If n\not\in\mathcal{O}_n, then \mathcal{O}_n=\mathcal{O}_{n-1}.

Proof idea: As n does not belong to \mathcal{O}_n  implies that \mathcal{O}_n needs to schedule 1,\dots,n-1.

For the more general case, let 1\le j\le n and consider the problem on jobs  1,\dots, j. Then by the same argument as we used to prove Claims 1 and 2, we have:

  • If j\in \mathcal{O}_j, then  \mathcal{O}_j = (\mathcal{O}_{p(n)}, j), i.e.   OPT(j)=v_j+OPT(p(j)).
  • If j\not\in \mathcal{O}_j, then  \mathcal{O}_j = \mathcal{O}_{j-1}, i.e. OPT(j)=OPT(j-1).

Thus, we have shown that

OPT(j) = \max\{v_j + OPT(p(j)), OPT(j-1)\}.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s


%d bloggers like this: