如何利用动态规划确定最长增长子序列？

Question

如何利用动态规划确定最长增长子序列？我有一组整数。我想找到最长增长子序列使用动态规划。

Answer 1

Petar Minchev的解释帮助我澄清了问题，但是我很难解析所有的东西，所以我使用了描述性很强的变量名和大量的注释来实现Python。我做了一个简单的递归解，O(n^2)解和O(N Log N)解。

我希望它能帮助清理算法！

递归解

def recursive_solution(remaining_sequence, bigger_than=None):
    """Finds the longest increasing subsequence of remaining_sequence that is      
    bigger than bigger_than and returns it.  This solution is O(2^n)."""

    # Base case: nothing is remaining.                                             
    if len(remaining_sequence) == 0:
        return remaining_sequence    # Recursive case 1: exclude the current element and process the remaining.     
    best_sequence = recursive_solution(remaining_sequence[1:], bigger_than)

    # Recursive case 2: include the current element if it's big enough.            
    first = remaining_sequence[0]

    if (first > bigger_than) or (bigger_than is None):

        sequence_with = [first] + recursive_solution(remaining_sequence[1:], first)

        # Choose whichever of case 1 and case 2 were longer.                         
        if len(sequence_with) >= len(best_sequence):
            best_sequence = sequence_with    return best_sequence

O(n^2)动态规划解

def dynamic_programming_solution(sequence):
    """Finds the longest increasing subsequence in sequence using dynamic          
    programming.  This solution is O(n^2)."""

    longest_subsequence_ending_with = []
    backreference_for_subsequence_ending_with = []
    current_best_end = 0

    for curr_elem in range(len(sequence)):
        # It's always possible to have a subsequence of length 1.                    
        longest_subsequence_ending_with.append(1)

        # If a subsequence is length 1, it doesn't have a backreference.             
        backreference_for_subsequence_ending_with.append(None)

        for prev_elem in range(curr_elem):
            subsequence_length_through_prev = (longest_subsequence_ending_with[prev_elem] + 1)

            # If the prev_elem is smaller than the current elem (so it's increasing)   
            # And if the longest subsequence from prev_elem would yield a better       
            # subsequence for curr_elem.                                               
            if ((sequence[prev_elem] < sequence[curr_elem]) and
                    (subsequence_length_through_prev >
                         longest_subsequence_ending_with[curr_elem])):

                # Set the candidate best subsequence at curr_elem to go through prev.    
                longest_subsequence_ending_with[curr_elem] = (subsequence_length_through_prev)
                backreference_for_subsequence_ending_with[curr_elem] = prev_elem                
                # If the new end is the best, update the best.    

        if (longest_subsequence_ending_with[curr_elem] >
                longest_subsequence_ending_with[current_best_end]):
            current_best_end = curr_elem            # Output the overall best by following the backreferences.  

    best_subsequence = []
    current_backreference = current_best_end    while current_backreference is not None:
        best_subsequence.append(sequence[current_backreference])
        current_backreference = (backreference_for_subsequence_ending_with[current_backreference])

    best_subsequence.reverse()

    return best_subsequence

O(N Log N)动态规划解

def find_smallest_elem_as_big_as(sequence, subsequence, elem):
    """Returns the index of the smallest element in subsequence as big as          
    sequence[elem].  sequence[elem] must not be larger than every element in       
    subsequence.  The elements in subsequence are indices in sequence.  Uses       
    binary search."""

    low = 0
    high = len(subsequence) - 1

    while high > low:
        mid = (high + low) / 2
        # If the current element is not as big as elem, throw out the low half of    
        # sequence.                                                                  
        if sequence[subsequence[mid]] < sequence[elem]:
            low = mid + 1
            # If the current element is as big as elem, throw out everything bigger, but 
        # keep the current element.                                                  
        else:
            high = mid    return highdef optimized_dynamic_programming_solution(sequence):
    """Finds the longest increasing subsequence in sequence using dynamic          
    programming and binary search (per                                             
    http://en.wikipedia.org/wiki/Longest_increasing_subsequence).  This solution   
    is O(n log n)."""

    # Both of these lists hold the indices of elements in sequence and not the        
    # elements themselves.                                                         
    # This list will always be sorted.                                             
    smallest_end_to_subsequence_of_length = []

    # This array goes along with sequence (not                                     
    # smallest_end_to_subsequence_of_length).  Following the corresponding element 
    # in this array repeatedly will generate the desired subsequence.              
    parent = [None for _ in sequence]

    for elem in range(len(sequence)):
        # We're iterating through sequence in order, so if elem is bigger than the   
        # end of longest current subsequence, we have a new longest increasing          
        # subsequence.                                                               
        if (len(smallest_end_to_subsequence_of_length) == 0 or
                    sequence[elem] > sequence[smallest_end_to_subsequence_of_length[-1]]):
            # If we are adding the first element, it has no parent.  Otherwise, we        
            # need to update the parent to be the previous biggest element.            
            if len(smallest_end_to_subsequence_of_length) > 0:
                parent[elem] = smallest_end_to_subsequence_of_length[-1]
            smallest_end_to_subsequence_of_length.append(elem)
        else:
            # If we can't make a longer subsequence, we might be able to make a        
            # subsequence of equal size to one of our earlier subsequences with a         
            # smaller ending number (which makes it easier to find a later number that 
            # is increasing).                                                          
            # Thus, we look for the smallest element in                                
            # smallest_end_to_subsequence_of_length that is at least as big as elem       
            # and replace it with elem.                                                
            # This preserves correctness because if there is a subsequence of length n 
            # that ends with a number smaller than elem, we could add elem on to the   
            # end of that subsequence to get a subsequence of length n+1.              
            location_to_replace = find_smallest_elem_as_big_as(sequence, smallest_end_to_subsequence_of_length, elem)
            smallest_end_to_subsequence_of_length[location_to_replace] = elem            
            # If we're replacing the first element, we don't need to update its parent 
            # because a subsequence of length 1 has no parent.  Otherwise, its parent  
            # is the subsequence one shorter, which we just added onto.                
            if location_to_replace != 0:
                parent[elem] = (smallest_end_to_subsequence_of_length[location_to_replace - 1])

    # Generate the longest increasing subsequence by backtracking through parent.  
    curr_parent = smallest_end_to_subsequence_of_length[-1]
    longest_increasing_subsequence = []

    while curr_parent is not None:
        longest_increasing_subsequence.append(sequence[curr_parent])
        curr_parent = parent[curr_parent]

    longest_increasing_subsequence.reverse()

    return longest_increasing_subsequence

Answer 2

谈到dp解决方案，我感到惊讶的是，没有人提到lis可以简化为LCS..您所需要做的就是对原始序列的副本进行排序，删除所有副本并对它们执行LCS。在伪码中是：

def LIS(S):
    T = sort(S)
    T = removeDuplicates(T)
    return LCS(S, T)

并以Go语言编写了完整的实现程序。如果不需要重构解决方案，则不需要维护整个n^2 dp矩阵。

func lcs(arr1 []int) int {
    arr2 := make([]int, len(arr1))
    for i, v := range arr1 {
        arr2[i] = v
    }
    sort.Ints(arr1)
    arr3 := []int{}
    prev := arr1[0] - 1
    for _, v := range arr1 {
        if v != prev {
            prev = v
            arr3 = append(arr3, v)
        }
    }

    n1, n2 := len(arr1), len(arr3)

    M := make([][]int, n2 + 1)
    e := make([]int, (n1 + 1) * (n2 + 1))
    for i := range M {
        M[i] = e[i * (n1 + 1):(i + 1) * (n1 + 1)]
    }

    for i := 1; i <= n2; i++ {
        for j := 1; j <= n1; j++ {
            if arr2[j - 1] == arr3[i - 1] {
                M[i][j] = M[i - 1][j - 1] + 1
            } else if M[i - 1][j] > M[i][j - 1] {
                M[i][j] = M[i - 1][j]
            } else {
                M[i][j] = M[i][j - 1]
            }
        }
    }

    return M[n2][n1]
}