【优化算法】Greedy Randomized Adaptive Search算法 超详细解析，附代码实现TSP问题求解

01 概述

Greedy Randomized Adaptive Search，贪婪随机自适应搜索（GRAS），是组合优化问题中的多起点元启发式算法，在算法的每次迭代中，主要由两个阶段组成：构造（construction）和局部搜索（ local search）。 构造（construction）阶段主要用于生成一个可行解，然后该初始可行解会被放进局部搜索进行邻域搜索，直到找到一个局部最优解为止。

02 总体框架

如上面所说，其实整一个算法的框架相对于其余算法来讲还算比较简单明了，你们能够先看如下总体的伪代码：

GRAS主要由两部分组成：

• Greedy_Randomized_Construction:在贪心的基础上，加入必定的随机因素，构造初始可行解。
• Local Search：对上面构造的初始可行解进行邻域搜索，直到找到一个局部最优解。

而后再多说两句：

1. Repair是什么鬼？
   有时候因为随机因素的加入，Greedy_Randomized_Construction阶段生成的解不必定都是可行解，因此为了保证下一步的Local Search能继续进行，加入repair算子，对解进行修复，保证其可行。

2. 不是说自适应（Adaptive）吗？我怎么没看到Adaptive 的过程？
   别急，这个后面具体举例的时候会详细讲到。

03 举个例子说明

为了你们能更加深刻理解该算法，咱们举一个简单的例子来为你们详细讲解算法的流程。

好了，相信你们都看懂上面的问题了（看不懂也别问我，摊手）。对于上述问题，咱们来一步一个脚印用GRAS来求解之，来，跟紧小编的脚步……

强调了不少次，GRAS由两部分组成：Greedy_Randomized_Construction和Local Search，因此，在求解具体问题的时候，完成这两部分的设计，而后按照第二节所示的框架搭起来就能够。

3.1 Greedy_Randomized_Construction

这里仍是老规矩，先上伪代码给你们看看，而后咱们再进行讲解，毕竟对于算法来讲，伪代码的做用不言而喻。

• 第1行，一开始解是一个空集。
• 第2行，初始化候选元素的集合，这里候选元素是指能放进Solution的元素（也就是目前Solution里面没有的），好比１，２，３……。
• 第3行，对候选集合的每一个元素进行评估，计算将元素x放入Solution会致使目标函数f改变量delta_x。
• 第5行，根据delta_x对各个元素排序，选取部分较好的候选元素组成RCL表（贪心性体如今这里）。
• 第6行，随机在RCL中选取一个元素放进Solution。（算法的随机性）
• 第八、9行，更新候选元素集合，而后对每一个元素进行从新评估计算delta值。（算法的自适应性体如今这里）

相信通过上面如此详细的介绍，你们都懂了吧！

3.2 Local Search

关于Local Search方面的内容，相信你们学习heuristic这么久了，就不用我多说什么了吧：

简单看一下伪代码便可，主要是邻域算子的设计，而后就是在邻域里面进行搜索，找到一个局部最优解为止。而后关于邻域搜索，有best-improving or first-improving strategy 两种策略，这个下次有时间出个专题给你们讲明白一些相关概念吧。

04 再论Greedy_Randomized_Construction

前面咱们说了，Greedy_Randomized_Construction用于生成初始解，既然是Greedy_Randomized两个结合体，那么确定就有一个权重分配的问题，即，是Greedy成分多一点呢？仍是Randomized成分多一点好呢？所以，为了控制这两个小老弟的权重，防止某个家伙在该过程当中用力过猛致使解不那么好的状况，咱们引入一个参数α：

其余部分就再也不多说，能够看到，上面的α参数主要是控制RCL的长度：

• 当α=0时，纯贪心，只能选取最优的候选元素。
• 当α=1时，纯随机，全部候选元素均可随机选。

05 代码实现

因为小编精力有限，就不从头写一遍了，从GitHub上找了一个感受还不错的算法给你们，也是求解TSP问题的。不过说实在的，python写算法的速度是很慢的，不管是速度仍是算法架构等方面都不推荐你们用matlab或者python写大型优化算法。

运行结果以下：

代码算例以及相关运行结果请关注公众号【程序猿声】，后台回复：GRAS，便可下载

############################################################################

# Created by: Prof. Valdecy Pereira, D.Sc.
# UFF - Universidade Federal Fluminense (Brazil)
# email:  valdecy.pereira@gmail.com
# Course: Metaheuristics
# Lesson: Local Search-GRASP

# Citation: 
# PEREIRA, V. (2018). Project: Metaheuristic-Local_Search-GRASP, File: Python-MH-Local Search-GRASP.py, GitHub repository: <https://github.com/Valdecy/Metaheuristic-Local_Search-GRASP>

############################################################################

# Required Libraries
import pandas as pd
import random
import numpy  as np
import copy
import os
from matplotlib import pyplot as plt 

# Function: Tour Distance
def distance_calc(Xdata, city_tour):
    distance = 0
    for k in range(0, len(city_tour[0])-1):
        m = k + 1
        distance = distance + Xdata.iloc[city_tour[0][k]-1, city_tour[0][m]-1]            
    return distance

# Function: Euclidean Distance 
def euclidean_distance(x, y):       
    distance = 0
    for j in range(0, len(x)):
        distance = (x.iloc[j] - y.iloc[j])**2 + distance   
    return distance**(1/2) 

# Function: Initial Seed
def seed_function(Xdata):
    seed = [[],float("inf")]
    sequence = random.sample(list(range(1,Xdata.shape[0]+1)), Xdata.shape[0])
    sequence.append(sequence[0])
    seed[0] = sequence
    seed[1] = distance_calc(Xdata, seed)
    return seed

# Function: Build Distance Matrix
def buid_distance_matrix(coordinates):
    Xdata = pd.DataFrame(np.zeros((coordinates.shape[0], coordinates.shape[0])))
    for i in range(0, Xdata.shape[0]):
        for j in range(0, Xdata.shape[1]):
            if (i != j):
                x = coordinates.iloc[i,:]
                y = coordinates.iloc[j,:]
                Xdata.iloc[i,j] = euclidean_distance(x, y)        
    return Xdata

# Function: Tour Plot
def plot_tour_distance_matrix (Xdata, city_tour):
    m = Xdata.copy(deep = True)
    for i in range(0, Xdata.shape[0]):
        for j in range(0, Xdata.shape[1]):
            m.iloc[i,j] = (1/2)*(Xdata.iloc[0,j]**2 + Xdata.iloc[i,0]**2 - Xdata.iloc[i,j]**2)    
    m = m.values
    w, u = np.linalg.eig(np.matmul(m.T, m))
    s = (np.diag(np.sort(w)[::-1]))**(1/2) 
    coordinates = np.matmul(u, s**(1/2))
    coordinates = coordinates.real[:,0:2]
    xy = pd.DataFrame(np.zeros((len(city_tour[0]), 2)))
    for i in range(0, len(city_tour[0])):
        if (i < len(city_tour[0])):
            xy.iloc[i, 0] = coordinates[city_tour[0][i]-1, 0]
            xy.iloc[i, 1] = coordinates[city_tour[0][i]-1, 1]
        else:
            xy.iloc[i, 0] = coordinates[city_tour[0][0]-1, 0]
            xy.iloc[i, 1] = coordinates[city_tour[0][0]-1, 1]
    plt.plot(xy.iloc[:,0], xy.iloc[:,1], marker = 's', alpha = 1, markersize = 7, color = 'black')
    plt.plot(xy.iloc[0,0], xy.iloc[0,1], marker = 's', alpha = 1, markersize = 7, color = 'red')
    plt.plot(xy.iloc[1,0], xy.iloc[1,1], marker = 's', alpha = 1, markersize = 7, color = 'orange')
    return

# Function: Tour Plot
def plot_tour_coordinates (coordinates, city_tour):
    coordinates = coordinates.values
    xy = pd.DataFrame(np.zeros((len(city_tour[0]), 2)))
    for i in range(0, len(city_tour[0])):
        if (i < len(city_tour[0])):
            xy.iloc[i, 0] = coordinates[city_tour[0][i]-1, 0]
            xy.iloc[i, 1] = coordinates[city_tour[0][i]-1, 1]
        else:
            xy.iloc[i, 0] = coordinates[city_tour[0][0]-1, 0]
            xy.iloc[i, 1] = coordinates[city_tour[0][0]-1, 1]
    plt.plot(xy.iloc[:,0], xy.iloc[:,1], marker = 's', alpha = 1, markersize = 7, color = 'black')
    plt.plot(xy.iloc[0,0], xy.iloc[0,1], marker = 's', alpha = 1, markersize = 7, color = 'red')
    plt.plot(xy.iloc[1,0], xy.iloc[1,1], marker = 's', alpha = 1, markersize = 7, color = 'orange')
    return

# Function: Rank Cities by Distance
def ranking(Xdata, city = 0):
    rank = pd.DataFrame(np.zeros((Xdata.shape[0], 2)), columns = ['Distance', 'City'])
    for i in range(0, rank.shape[0]):
        rank.iloc[i,0] = Xdata.iloc[i,city]
        rank.iloc[i,1] = i + 1
    rank = rank.sort_values(by = 'Distance')
    return rank

# Function: RCL
def restricted_candidate_list(Xdata, greediness_value = 0.5):
    seed = [[],float("inf")]
    sequence = []
    sequence.append(random.sample(list(range(1,Xdata.shape[0]+1)), 1)[0])
    for i in range(0, Xdata.shape[0]):
        count = 1
        rand = int.from_bytes(os.urandom(8), byteorder = "big") / ((1 << 64) - 1)
        if (rand > greediness_value and len(sequence) < Xdata.shape[0]):
            next_city = int(ranking(Xdata, city = sequence[-1] - 1).iloc[count,1])
            while next_city in sequence:
                count = np.clip(count+1,1,Xdata.shape[0]-1)
                next_city = int(ranking(Xdata, city = sequence[-1] - 1).iloc[count,1])
            sequence.append(next_city)
        elif (rand <= greediness_value and len(sequence) < Xdata.shape[0]):
            next_city = random.sample(list(range(1,Xdata.shape[0]+1)), 1)[0]
            while next_city in sequence:
                next_city = int(random.sample(list(range(1,Xdata.shape[0]+1)), 1)[0])
            sequence.append(next_city)
    sequence.append(sequence[0])
    seed[0] = sequence
    seed[1] = distance_calc(Xdata, seed)
    return seed

# Function: 2_opt
def local_search_2_opt(Xdata, city_tour):
    tour = copy.deepcopy(city_tour)
    best_route = copy.deepcopy(tour)
    seed = copy.deepcopy(tour)  
    for i in range(0, len(tour[0]) - 2):
        for j in range(i+1, len(tour[0]) - 1):
            best_route[0][i:j+1] = list(reversed(best_route[0][i:j+1]))           
            best_route[0][-1]  = best_route[0][0]                          
            best_route[1] = distance_calc(Xdata, best_route)           
            if (best_route[1] < tour[1]):
                tour[1] = copy.deepcopy(best_route[1])
                for n in range(0, len(tour[0])): 
                    tour[0][n] = best_route[0][n]          
            best_route = copy.deepcopy(seed) 
    return tour

# Function: GRASP
def greedy_randomized_adaptive_search_procedure(Xdata, city_tour, iterations = 50, rcl = 25, greediness_value = 0.5):
    count = 0
    best_solution = copy.deepcopy(city_tour)
    while (count < iterations):
        rcl_list = []
        for i in range(0, rcl):
            rcl_list.append(restricted_candidate_list(Xdata, greediness_value = greediness_value))
        candidate = int(random.sample(list(range(0,rcl)), 1)[0])
        city_tour = local_search_2_opt(Xdata, city_tour = rcl_list[candidate])
        while (city_tour[0] != rcl_list[candidate][0]):
            rcl_list[candidate] = copy.deepcopy(city_tour)
            city_tour = local_search_2_opt(Xdata, city_tour = rcl_list[candidate])
        if (city_tour[1] < best_solution[1]):
            best_solution = copy.deepcopy(city_tour) 
        count = count + 1
        print("Iteration =", count, "-> Distance =", best_solution[1])
    print("Best Solution =", best_solution)
    return best_solution

######################## Part 1 - Usage ####################################

X = pd.read_csv('Python-MH-Local Search-GRASP-Dataset-01.txt', sep = '\t') #17 cities = 1922.33
seed = seed_function(X)
lsgrasp = greedy_randomized_adaptive_search_procedure(X, city_tour = seed, iterations = 5, rcl = 5, greediness_value = 0.5)
plot_tour_distance_matrix(X, lsgrasp) # Red Point = Initial city; Orange Point = Second City # The generated coordinates (2D projection) are aproximated, depending on the data, the optimum tour may present crosses.

Y = pd.read_csv('Python-MH-Local Search-GRASP-Dataset-02.txt', sep = '\t') # Berlin 52 = 7544.37
X = buid_distance_matrix(Y)
seed = seed_function(X)
lsgrasp = greedy_randomized_adaptive_search_procedure(X, city_tour = seed, iterations = 10, rcl = 15, greediness_value = 0.5)
plot_tour_coordinates (Y, lsgrasp) # Red Point = Initial city; Orange Point = Second City