dynamic programming_矩阵链乘法(matrix_chain_order)_python

theory

记住,矩阵子链$A_i\cdots A_k$链内各个矩阵相乘的结果矩阵的规模可以仅从该子链的第一个矩阵$A_i$的行数$p_{i-1}$和最后一个矩阵$A_k$的列数$p_k$有关
即,结果矩阵的规模为$p_{i-1}\times p_k$
这是理解$m[2,2]+m[3,5]+p_1{p}_2{p}_5$中为什么是$+p_1{p}_2{p}_5$的基础
$p_i为第i个矩阵的列数(i>=1),至于行数,则为p_{i-1}$
m[2,2]相当于单个矩阵构成的矩阵链(k=i=2),行数为$p_{2-1}=p_1(i=2)$,列数为$p_2(k=i=2)$,则,规格为$p_1\times p_2$

specific examples:

Construct the optimal solution

the parentheses solution:

python:

import math
def matrix_chain_order(list_p):
    """[summary]

    Args:
        list_p ([type]): [description]

    Returns:
        [type]: [description]
    """
    n = len(list_p)-1
    ''' we could think that the table_min_costs contains the optimal cost values of the different subproblem(sub_matrix chain) scales
    (from bottom to top to solve the problem)
    '''
    # the table to save the lowest costs to multiple the matrix chain
    table_min_costs = [[0 for i in range(0, n+1)]
                       for j in range(0, n+1)]  # (n+1)*(n+1)
    # the table_save_partition save the the optimal partition point:make the multiplication cost lowest each scale cases
    table_save_partition = [
        [0 for i in range(0, n+1)] for j in range(0, n+1)]  # (n)*(n)
    ''' the length of the sub_matrix chain is 1 cases could be centralized process:assign them as 0 in the table_min_costs '''
    for i in range(1, n+1):
        table_min_costs[i][i] = 0
    ''' the essential part of the algorithm: '''
    # length is the chain length(traverse the all kinds of sub matrix chains cases(length>1))
    for length in range(2, n+1):
        ''' each specific length cases have different partition schemes: 
        traverse all these divide cases:'''
        for start_i in range(1, n-length+2):  # the start_i is the index of each sub_matrix chain
            # make the j-i+1==l(namely ,the length of sub_matrix chain)
            end_i = start_i+length-1
            # initial the costs as infinite value:
            table_min_costs[start_i][end_i] = math.inf
            """ 
            # focus on the each sepecified sub_matrix chain
            # the following loop will test(run) j-i times:to find the optimal split point:
            # the index to_opt_partition is the index of the optimal split point(the matrix to_opt_partition matrix will be belong to the former subproble(sub_matrix chain)) 
            """
            for to_opt_partition in range(start_i, end_i):
                ''' the to_opt_partition >=start_i>=1 '''
                to_lowest_cost = table_min_costs[start_i][to_opt_partition] + table_min_costs[to_opt_partition+1][end_i] + \
                    list_p[start_i-1]*list_p[to_opt_partition]*list_p[end_i]
                if to_lowest_cost < table_min_costs[start_i][end_i]:
                    table_min_costs[start_i][end_i] = to_lowest_cost

                    table_save_partition[start_i][end_i] = to_opt_partition

    return table_min_costs, table_save_partition


def print_optimal_parentheses(s, i, j):
    """
    matrix_index=1
    Args:
        s ([list]): [table_save_partition(optimal)]
        i ([int]): [start_i index of the subproblem]
        j ([int]): [end_i index of the subproblem]
    """
    """ the simplest case (the sub_matrix chain length is 1)
    the case as the recursive exit:"""
    if i == j:
        print("A"+str(i), end="")
    else:
        print("(", end="")
        print_optimal_parentheses(s, i, s[i][j])
        print_optimal_parentheses(s, s[i][j]+1, j)
        print(")", end="")


def test(list_p):
    table_min_costs, table_save_partition = matrix_chain_order(list_p)
    print_optimal_parentheses(table_save_partition, 1, len(list_p)-1)


list_p = [30, 35, 15, 5, 10, 20, 25]


test(list_p)