Dynamic Programming. Shuang Zhao. Microsoft Research Asia September 5, Dynamic Programming. Shuang Zhao. Outline. Introduction.

Transcription

1 Microsoft Research Asia September 5, 2005

2

3 Section I

4 What is? Definition is a technique for efficiently recurrence computing by storing partial results. In this slides, I will NOT use too many formal words, but only look on some interesting problems.

5 The Longest Ascending Subsequence P : a 1, a 2,..., a n. Q : a b1, a b2,..., a bk, satisfying and 1 b 1 < b 2 <... < b k n a b1 < a b2 <... < a bk We say Q is an ascending subsequence of P. Your task is: given a sequence P, find the length its longest ascending subsequence (LAS).

6 The Longest Ascending Subsequence We use f i to denote the length of P s LAS ending with element a i. Let a 0 :=, f 0 := 0. Then we have f k = max 0 i<k {f i + 1 : a i < a k }, Hence the length of P s LAS is max{f i }. 1 k n This naive algorithm runs in O(n 2 ) time, and later we will look on this problem again.

7 The Shortest Hamilton Path Given a connected graph G(V, E) and a vertex s V. Given a weight function w over E, denoting the lengths of edges. Your task is to find the length of shortest Hamilton path starting from s.

8 The Shortest Hamilton Path This problem is N P-hard. When V is small, one FAST algorithm to solve this problem is. Use f i,s where i V, S S to denote the length of shortest Hamilton path over S, which ending at vertex i. Then min i V { fi,s } is what we want, and f i,s = min j S { fj,s {i} + w(j, i) } This algorithm runs in O( V 2 2 V ) time.

9 Section II

10 Problem I

11 The Problem: There is a board with N rows and 6 columns. How many ways can we cover the board with L pieces? One way to solve this problem, is. But where are the partial results?

12 This is a pattern, say P n,(101011)2 :

13 And by completely covering the last row of P n,(101011)2, we can obtain P n+1,(110110)2 and P n+1,(111100)2.

14 Observation For any pattern P i,j, after covering its (i + 1)-th row, we can only get patterns with the form of P i+1,j. We call P i+1,j can be generated from P i,j, denoting as P i+1,j P i,j.

15 Observation For any given covering of the N 6 board and an integer i (0 i < N), there will be one and only one pattern P i,j, which is contained in the covering. i := 4

16 The partial results are: Let f i,j to be the number of ways to cover pattern P i,j. And f i,j = f i,j = { 1 j = 0 0 otherwise P i,j P i 1,j f i 1,j i = 0 0 < i < n Finally f n 1,(111111)2 is the answer to this problem. By a coarse calculation, we know this algorithm runs in O(2 6 N 4 6 ) = O(N) time.

17 Problem II

18 The Problem: Given an alphabet Σ, a set of strings S f Σ, and an integer n. Your task is to count the number of n-length-strings which contain at least one string in S as its substring.

19 Let Σ := {a, b, c}, S := {ab} and n := 3. Then we have: aab aba abb abc bab cab One method to solve this problem is using the inclusion and exclusion theorem. But the time complexity of this method is quite high (at least O(2 S )).

20 Definition Let and be two binary relations over Σ : x y x is a prefix of y x y x is a suffix of y Definition We define the prefix set of S (S ) by pre(s) := { s : ( s S)(s s) } Obviously, S pre(s).

21 For instance, let Σ := {a, b, c} and S := {aa, ba, cba}, then pre(s) = {ε, a, aa, b, ba, c, cb, cba}

22 Let pre(s) = {s 0, s 1,..., s t } where s 0 = ε. Let F i (i = 0, 1,..., t) be the subset of Σ, satisfying for all s F i : s i s ( s s )[( r S)(r s )] s j (s i s j s j s )

23 Observation Given p Σ, for all s F i, there will be exact one j (0 j t) satisfying s p F j. We say F i F j.

24 Define f i,j by the number of i-length strings in F j. The we have f i,j = F j F k f i 1,k and the number of strings which contain at least one string in S as their substring is n f i,j Σ n i i=1 s j S

25 A brute-force approach to determine the binary relation takes O( pre(s) Σ L) O(L 2 ) where L is the total length of strings in pre(s). The time complexity of doing the is O(n pre(s) Σ ) O(n pre(s) ) O(n L).

26 Section III

27 Speed up partial results calculation

28 Speed up partial results calculation Longest Ascending Subsequence (LAS) problem revisited: The naive algorithm runs in O(n 2 ) time. Can we solve this problem more efficiently?

29 Speed up partial results calculation Let n := 6, {a n } := {5, 1, 6, 2, 4, 1.5}: Vertices, Contours Red 1, Green 2, Blue 3, etc.

30 Speed up partial results calculation Proposition The contours will NOT intersect. Proof. Assume two contours with levels i and j (i < j) intersects. Then it holds that at least one vertex of level j is below contours i. This gives the contradiction that one vertex of level i is below contours i.

31 Speed up partial results calculation {a n } = {5}

32 Speed up partial results calculation {a n } = {5, 1}

33 Speed up partial results calculation {a n } = {5, 1, 6}

34 Speed up partial results calculation {a n } = {5, 1, 6, 2}

35 Speed up partial results calculation {a n } = {5, 1, 6, 2, 4}

36 Speed up partial results calculation {a n } = {5, 1, 6, 2, 4, 1.5}

37 Speed up partial results calculation How to update the contours?

38 Speed up partial results calculation How to update the contours? 1 When the new vertex is above all contours Then a new contour is created.

39 Speed up partial results calculation How to update the contours? 2 When the new vertex is below some contour Then the contour immediately above the new vertex, is lowered.

40 Speed up partial results calculation How to update the contours? 3 When the new vertex is just on some contour Then the contours remain the same.

41 Speed up partial results calculation Find the contour immediately above the new vertex A brute-force approach runs in O(n) time. A Binary Search algorithm takes only O(log n) time. So by using Binary Search, the time complexity of entry algorithm is reduced to O(n log n).

42 Speed up partial results calculation Further thinking How to implement this algorithm? How to find the longest non-descending subsequence of a given sequence? If we given a weight to every number in the sequence, how to find the ascending subsequence which maximizes the sum of weights?

43 Speed up by monotonicity

44 Speed up by monotonicity The Optimal Binary Search Tree (OBST) Problem n numbers a 1 < a 2 <... < a n n weights w 1, w 2,..., w n 0 Construct a binary search tree using a 1,..., a n. Cost = n w i d i i=1 Your task is to find a binary search tree which minimizes the cost.

45 Speed up by monotonicity Let n := 5 and {w n } := {1, 2, 2, 3, 1}. The cost of OBST is 18.

46 Speed up by monotonicity The naive approach f i,j : The cost of OBST constructed by w i, w i+1,..., w j Then f i,j = 0 (i > j) f i,j = min {f i,k 1 + f k+1,j } + w t (i j) i k j The answer is f 1,n. The time complexity is O(n 3 ). How to speed up this algorithm? i t j

47 Speed up by monotonicity Definition For a given m n matrix A, if for all i 1 i 2 j 1 j 2, it holds A[i 1, j 1 ] + A[i 2, j 2 ] A[i 1, j 2 ] + A[i 2, j 1 ] then we say A is totally monotonic.

48 Speed up by monotonicity The inequality A[i 1, j 1 ] + A[i 2, j 2 ] A[i 1, j 2 ] + A[i 2, j 1 ] is called Quadrangle Inequality. j j 2 A 12 A 22 j 1 A 11 A 21 i 1 i 2 i

49 Speed up by monotonicity Recurrent formula of OBST problem f i,j = 0 (i > j) f i,j = min {f i,k 1 + f k+1,j } + w t (i j) i k j Define F, W R n n + by Then when i j F [i, j] := f i,j i t j W [i, j] := i k j w k F [i, j] = min {F [i, k 1] + F [k + 1, j]} + W [i, j] i k j

50 Speed up by monotonicity Proposition Matrix W is totally monotonic. Proof. For all i 1 i 2 < j 1 j 2, W [i 1, j 1 ] + W [i 2, j 2 ] = W [i 1, j 2 ] + W [i 2, j 1 ] i 1 i 2 j 1 j 2

51 Speed up by monotonicity Proposition Matrix F is also totally monotonic. This can be proved by induction on j 2 i 1. To see the details, read in the section.

52 Speed up by monotonicity Definition For all 1 i j n, define s(i, j) by s(i, j) := max i k j { k : F [i, j] = F [i, k 1] + F [k + 1, j] + W [i, j] }

53 Speed up by monotonicity Proposition s(i, j) is monotonic, namely for all 1 i j < n, s(i, j) s(i, j + 1) s(i + 1, j + 1) To see the proof, read in the section.

54 Speed up by monotonicity The new recurrent formula f i,j = min {f i,k 1 + f k+1,j } + s(i,j 1) k s(i+1,j) The time complexity to solve it is O(n 2 ). i t j w t (i j)

55

56 (In the OBST problem) Proposition Matrix F is also totally monotonic. Proof (Part 1). When i 1 = i 2 or j 1 = j 2 the Quadrangle Inequality holds.

57 Proof (Part 2). When i 1 < i 2 = j 1 < j 2, we prove by induction on j 2 i 1. Let F [i 1, j 2 ] = F [i 1, k 1] + F [k + 1, j 2 ] + W [i 1, j 2 ] Without loss of generality, k j 1.

58 Case 1. k < j 1 (= i 2 ) F [i 1, j 1 ] + F [i 2, j 2 ] F [i 1, k 1] + F [k + 1, j 1 ] + W [i 1, j 1 ] + F [i 2, j 2 ] F [i 1, k 1] + W [i 1, j 1 ] + F [k + 1, j 2 ] + F [i 2, j 1 ] F [i 1, k 1] + W [i 1, j 2 ] + F [k + 1, j 2 ] + F [i 2, j 1 ] = F [i 1, j 2 ] + F [i 2, j 1 ]

59 Case 2. k = j 1 (= i 2 ) F [i 1, j 1 ] + F [i 2, j 2 ] F [i 1, k 1] + W [i 1, j 1 ] + F [i 2, j 2 ] F [i 1, k 1] + W [i 1, j 1 ] + F [j 1 + 1, j 2 ] + W [i 2, j 2 ] = F [i 1, k 1] + W [i 1, j 2 ] + F [k + 1, j 2 ] + W [i 2, j 1 ] = F [i 1, j 2 ] + F [i 2, j 1 ]

60 Proof (Part 3). When i 1 < i 2 < j 1 < j 2, we prove by induction on j 2 i 1. Let F [i 2, j 1 ] = F [i 2, k 1] + F [k + 1, j 1 ] + W [i 2, j 1 ] F [i 1, j 2 ] = F [i 1, t 1] + F [t + 1, j 2 ] + W [i 1, j 2 ] Without loss of generality, t k. Hence i 1 t k j 1.

61 Then we have F [i 1, j 1 ] + F [i 2, j 2 ] F [i 1, t 1] + F [t + 1, j 1 ] + W [i 1, j 1 ] + F [i 2, k 1] + F [k + 1, j 2 ] + W [i 2, j 2 ] F [i 1, t 1] + F [t + 1, j 2 ] + W [i 1, j 2 ] + F [i 2, k 1] + F [k + 1, j 1 ] + W [i 2, j 1 ] = F [i 1, j 2 ] + F [i 2, j 1 ]

62 (In the OBST problem) Proposition s(i, j) is monotonic, namely for all 1 i j < n, s(i, j) s(i, j + 1) s(i + 1, j + 1) Proof. By symmetry, we need only to prove that s(i, j) s(i, j + 1). When i = j, s(i, j) = i s(i, j + 1).

63 Next we assume i < j. For convenience, we use symbol F k [i, j] to be the short from of F [i, k 1] + F [k + 1, j] + W [i, j]. So F s(i,j) [i, j] = F [i, j]. Since matrix F is totally monotonic, for all k k j, F [k + 1, j] + F [k + 1, j + 1] F [k + 1, j] + F [k + 1, j + 1]

64 Thus F [i, k 1] + F [k + 1, j] + W [i, j] + F [i, k 1] + F [k + 1, j + 1] + W [i, j + 1] F [i, k 1] + F [k + 1, j + 1] + W [i, j + 1] + F [i, k 1] + F [k + 1, j] + W [i, j] namely F k [i, j] + F k [i, j + 1] F k [i, j + 1] + F k [i, j] that is F k [i, j] F k [i, j] F k [i, j + 1] F k [i, j + 1]

65 Therefore F k [i, j] F k [i, j] F k [i, j + 1] F k [i, j + 1] For all k < s(i, j), F s(i,j) [i, j] = F [i, j] F k [i, j]. So F s(i,j) (i, j + 1) F k (i, j + 1). Hence F s(i,j) [i, j + 1] F k [i, j + 1]. This gives s(i, j) s(i, j + 1).