Hungarian Algorithm for Assignment Problem | Set 1 (Introduction)
Let there be n agents and n tasks. Any agent can be assigned to perform any task, incurring some cost that may vary depending on the agent-task assignment. It is required to perform all tasks by assigning exactly one agent to each task and exactly one task to each agent in such a way that the total cost of the assignment is minimized. Example: You work as a manager for a chip manufacturer, and you currently have 3 people on the road meeting clients. Your salespeople are in Jaipur, Pune and Bangalore, and you want them to fly to three other cities: Delhi, Mumbai and Kerala. The table below shows the cost of airline tickets in INR between the cities: 
The question: where would you send each of your salespeople in order to minimize fair? Possible assignment: Cost = 11000 INR 
Other Possible assignment: Cost = 9500 INR and this is the best of the 3! possible assignments. 
Brute force solution is to consider every possible assignment implies a complexity of Ω(n!). The Hungarian algorithm, aka Munkres assignment algorithm, utilizes the following theorem for polynomial runtime complexity (worst case O(n3)) and guaranteed optimality: If a number is added to or subtracted from all of the entries of any one row or column of a cost matrix, then an optimal assignment for the resulting cost matrix is also an optimal assignment for the original cost matrix. We reduce our original weight matrix to contain zeros, by using the above theorem. We try to assign tasks to agents such that each agent is doing only one task and the penalty incurred in each case is zero. Core of the algorithm (assuming square matrix):
- For each row of the matrix, find the smallest element and subtract it from every element in its row.
- Do the same (as step 1) for all columns.
- Cover all zeros in the matrix using minimum number of horizontal and vertical lines.
- Test for Optimality: If the minimum number of covering lines is n, an optimal assignment is possible and we are finished. Else if lines are lesser than n, we haven’t found the optimal assignment, and must proceed to step 5.
- Determine the smallest entry not covered by any line. Subtract this entry from each uncovered row, and then add it to each covered column. Return to step 3.
Explanation for above simple example:
Below is the cost matrix of example given in above diagrams. 2500 4000 3500 4000 6000 3500 2000 4000 2500 Step 1: Subtract minimum of every row. 2500, 3500 and 2000 are subtracted from rows 1, 2 and 3 respectively. 0 1500 1000 500 2500 0 0 2000 500 Step 2: Subtract minimum of every column. 0, 1500 and 0 are subtracted from columns 1, 2 and 3 respectively. 0 0 1000 500 1000 0 0 500 500 Step 3: Cover all zeroes with minimum number of horizontal and vertical lines. Step 4: Since we need 3 lines to cover all zeroes, we have found the optimal assignment. 2500 4000 3500 4000 6000 3500 2000 4000 2500 So the optimal cost is 4000 + 3500 + 2000 = 9500
An example that doesn’t lead to optimal value in first attempt: In the above example, the first check for optimality did give us solution. What if we the number covering lines is less than n.
cost matrix: 1500 4000 4500 2000 6000 3500 2000 4000 2500 Step 1: Subtract minimum of every row. 1500, 2000 and 2000 are subtracted from rows 1, 2 and 3 respectively. 0 2500 3000 0 4000 1500 0 2000 500 Step 2: Subtract minimum of every column. 0, 2000 and 500 are subtracted from columns 1, 2 and 3 respectively. 0 500 2500 0 2000 1000 0 0 0 Step 3: Cover all zeroes with minimum number of horizontal and vertical lines. Step 4: Since we only need 2 lines to cover all zeroes, we have NOT found the optimal assignment. Step 5: We subtract the smallest uncovered entry from all uncovered rows. Smallest entry is 500. -500 0 2000 -500 1500 500 0 0 0 Then we add the smallest entry to all covered columns, we get 0 0 2000 0 1500 500 500 0 0 Now we return to Step 3:. Here we cover again using lines. and go to Step 4:. Since we need 3 lines to cover, we found the optimal solution. 1500 4000 4500 2000 6000 3500 2000 4000 2500 So the optimal cost is 4000 + 2000 + 2500 = 8500
C++
#include<bits/stdc++.h>using namespace std;class Solution { public: int cost[31][31]; //cost matrix int n, max_match; //n workers and n jobs int lx[31], ly[31]; //labels of X and Y parts int xy[31]; //xy[x] - vertex that is matched with x, int yx[31]; //yx[y] - vertex that is matched with y bool S[31], T[31]; //sets S and T in algorithm int slack[31]; //as in the algorithm description int slackx[31]; //slackx[y] such a vertex, that int prev_ious[31]; //array for memorizing alternating p void init_labels() { memset(lx, 0, sizeof(lx)); memset(ly, 0, sizeof(ly)); for (int x = 0; x < n; x++) for (int y = 0; y < n; y++) lx[x] = max(lx[x], cost[x][y]); } void update_labels() { int x, y; int delta = 99999999; //init delta as infinity for (y = 0; y < n; y++) //calculate delta using slack if (!T[y]) delta = min(delta, slack[y]); for (x = 0; x < n; x++) //update X labels if (S[x]) lx[x] -= delta; for (y = 0; y < n; y++) //update Y labels if (T[y]) ly[y] += delta; for (y = 0; y < n; y++) //update slack array if (!T[y]) slack[y] -= delta; } void add_to_tree(int x, int prev_iousx) //x - current vertex,prev_iousx - vertex from X before x in the alternating path, //so we add edges (prev_iousx, xy[x]), (xy[x], x) { S[x] = true; //add x to S prev_ious[x] = prev_iousx; //we need this when augmenting for (int y = 0; y < n; y++) //update slacks, because we add new vertex to S if (lx[x] + ly[y] - cost[x][y] < slack[y]) { slack[y] = lx[x] + ly[y] - cost[x][y]; slackx[y] = x; } } void augment() //main function of the algorithm { if (max_match == n) return; //check whether matching is already perfect int x, y, root; //just counters and root vertex int q[31], wr = 0, rd = 0; //q - queue for bfs, wr,rd - write and read //pos in queue memset(S, false, sizeof(S)); //init set S memset(T, false, sizeof(T)); //init set T memset(prev_ious, -1, sizeof(prev_ious)); //init set prev_ious - for the alternating tree for (x = 0; x < n; x++) //finding root of the tree { if (xy[x] == -1) { q[wr++] = root = x; prev_ious[x] = -2; S[x] = true; break; } } for (y = 0; y < n; y++) //initializing slack array { slack[y] = lx[root] + ly[y] - cost[root][y]; slackx[y] = root; } //second part of augment() function while (true) //main cycle { while (rd < wr) //building tree with bfs cycle { x = q[rd++]; //current vertex from X part for (y = 0; y < n; y++) //iterate through all edges in equality graph if (cost[x][y] == lx[x] + ly[y] && !T[y]) { if (yx[y] == -1) break; //an exposed vertex in Y found, so //augmenting path exists! T[y] = true; //else just add y to T, q[wr++] = yx[y]; //add vertex yx[y], which is matched //with y, to the queue add_to_tree(yx[y], x); //add edges (x,y) and (y,yx[y]) to the tree } if (y < n) break; //augmenting path found! } if (y < n) break; //augmenting path found! update_labels(); //augmenting path not found, so improve labeling wr = rd = 0; for (y = 0; y < n; y++) //in this cycle we add edges that were added to the equality graph as a //result of improving the labeling, we add edge (slackx[y], y) to the tree if //and only if !T[y] && slack[y] == 0, also with this edge we add another one //(y, yx[y]) or augment the matching, if y was exposed if (!T[y] && slack[y] == 0) { if (yx[y] == -1) //exposed vertex in Y found - augmenting path exists! { x = slackx[y]; break; } else { T[y] = true; //else just add y to T, if (!S[yx[y]]) { q[wr++] = yx[y]; //add vertex yx[y], which is matched with //y, to the queue add_to_tree(yx[y], slackx[y]); //and add edges (x,y) and (y, //yx[y]) to the tree } } } if (y < n) break; //augmenting path found! } if (y < n) //we found augmenting path! { max_match++; //increment matching //in this cycle we inverse edges along augmenting path for (int cx = x, cy = y, ty; cx != -2; cx = prev_ious[cx], cy = ty) { ty = xy[cx]; yx[cy] = cx; xy[cx] = cy; } augment(); //recall function, go to step 1 of the algorithm } }//end of augment() function int hungarian() { int ret = 0; //weight of the optimal matching max_match = 0; //number of vertices in current matching memset(xy, -1, sizeof(xy)); memset(yx, -1, sizeof(yx)); init_labels(); //step 0 augment(); //steps 1-3 for (int x = 0; x < n; x++) //forming answer there ret += cost[x][xy[x]]; return ret; } int assignmentProblem(int Arr[], int N) { n = N; for(int i=0; i<n; i++) for(int j=0; j<n; j++) cost[i][j] = -1*Arr[i*n+j]; int ans = -1 * hungarian(); return ans; }};int main(){ int n=3; int Arr[3*3]={1500,4000,4500,2000,6000,3500,2000,4000,2500}; /*1500 4000 4500 2000 6000 3500 2000 4000 2500*/ Solution ob; cout<<ob.assignmentProblem(Arr,n)<<endl;} |
Java
// Java Equivalent:import java.util.*;public class Solution { public static int n; public static int cost[][]; //cost matrix public static int max_match; //n workers and n jobs public static int lx[], ly[]; //labels of X and Y parts public static int xy[]; //xy[x] - vertex that is matched with x, public static int yx[]; //yx[y] - vertex that is matched with y public static boolean S[], T[]; //sets S and T in algorithm public static int slack[]; //as in the algorithm description public static int slackx[]; //slackx[y] such a vertex, that public static int prev_ious[]; //array for memorizing alternating p public static void init_labels() { Arrays.fill(lx, 0); Arrays.fill(ly, 0); for (int x = 0; x < n; x++) for (int y = 0; y < n; y++) lx[x] = Math.max(lx[x], cost[x][y]); } public static void update_labels() { int x, y; int delta = 99999999; //init delta as infinity for (y = 0; y < n; y++) //calculate delta using slack if (!T[y]) delta = Math.min(delta, slack[y]); for (x = 0; x < n; x++) //update X labels if (S[x]) lx[x] -= delta; for (y = 0; y < n; y++) //update Y labels if (T[y]) ly[y] += delta; for (y = 0; y < n; y++) //update slack array if (!T[y]) slack[y] -= delta; } public static void add_to_tree(int x, int prev_iousx) //x - current vertex,prev_iousx - vertex from X before x in the alternating path, //so we add edges (prev_iousx, xy[x]), (xy[x], x) { S[x] = true; //add x to S prev_ious[x] = prev_iousx; //we need this when augmenting for (int y = 0; y < n; y++) //update slacks, because we add new vertex to S if (lx[x] + ly[y] - cost[x][y] < slack[y]) { slack[y] = lx[x] + ly[y] - cost[x][y]; slackx[y] = x; } } public static void augment() //main function of the algorithm { if (max_match == n) return; //check whether matching is already perfect int x, y; //just counters and root vertex int q[] = new int[n], wr = 0, rd = 0; //q - queue for bfs, wr,rd - write and read //pos in queue Arrays.fill(S, false); //init set S Arrays.fill(T, false); //init set T Arrays.fill(prev_ious, -1); //init set prev_ious - for the alternating tree int root = -1; for (x = 0; x < n; x++) //finding root of the tree { if (xy[x] == -1) { q[wr++] = root = x; prev_ious[x] = -2; S[x] = true; break; } } if (root == -1) { // All vertices are already matched return; } for (y = 0; y < n; y++) //initializing slack array { slack[y] = lx[root] + ly[y] - cost[root][y]; slackx[y] = root; } //second part of augment() function while (true) //main cycle { while (rd < wr) //building tree with bfs cycle { x = q[rd++]; //current vertex from X part for (y = 0; y < n; y++) //iterate through all edges in equality graph if (cost[x][y] == lx[x] + ly[y] && !T[y]) { if (yx[y] == -1) break; //an exposed vertex in Y found, so //augmenting path exists! T[y] = true; //else just add y to T, q[wr++] = yx[y]; //add vertex yx[y], which is matched //with y, to the queue add_to_tree(yx[y], x); //add edges (x,y) and (y,yx[y]) to the tree } if (y < n) break; //augmenting path found! } if (y < n) break; //augmenting path found! update_labels(); //augmenting path not found, so improve labeling wr = rd = 0; for (y = 0; y < n; y++) //in this cycle we add edges that were added to the equality graph as a //result of improving the labeling, we add edge (slackx[y], y) to the tree if //and only if !T[y] && slack[y] == 0, also with this edge we add another one //(y, yx[y]) or augment the matching, if y was exposed if (!T[y] && slack[y] == 0) { if (yx[y] == -1) //exposed vertex in Y found - augmenting path exists! { x = slackx[y]; break; } else { T[y] = true; //else just add y to T, if (!S[yx[y]]) { q[wr++] = yx[y]; //add vertex yx[y], which is matched with //y, to the queue add_to_tree(yx[y], slackx[y]); //and add edges (x,y) and (y, //yx[y]) to the tree } } } if (y < n) break; //augmenting path found! } if (y < n) //we found augmenting path! { max_match++; //increment matching //in this cycle we inverse edges along augmenting path for (int cx = x, cy = y, ty; cx != -2; cx = prev_ious[cx], cy = ty) { ty = xy[cx]; yx[cy] = cx; xy[cx] = cy; } augment(); //recall function, go to step 1 of the algorithm } }//end of augment() function public static int hungarian() { int ret = 0; //weight of the optimal matching max_match = 0; //number of vertices in current matching xy = new int[n]; yx = new int[n]; Arrays.fill(xy, -1); Arrays.fill(yx, -1); init_labels(); //step 0 augment(); //steps 1-3 for (int x = 0; x < n; x++) //forming answer there ret += cost[x][xy[x]]; return ret; } public static int assignmentProblem(int Arr[], int N) { n = N; cost = new int[n][n]; lx = new int[n]; ly = new int[n]; S = new boolean[n]; T = new boolean[n]; slack = new int[n]; slackx = new int[n]; prev_ious = new int[n]; for(int i=0; i<n; i++) for(int j=0; j<n; j++) cost[i][j] = -1*Arr[i*n+j]; int ans = -1 * hungarian(); return ans; } public static void main(String[] args) { int n=3; int Arr[]={1500,4000,4500,2000,6000,3500,2000,4000,2500}; /*1500 4000 4500 2000 6000 3500 2000 4000 2500*/ Solution ob = new Solution(); System.out.println(ob.assignmentProblem(Arr,n)); }} |
Python3
# python implementation of the above approach.class Solution: def __init__(self): # cost matrix self.cost = [[]]*31 for i in range(31): self.cost[i] = [0]*31 self.n = 0; # n workers and n jobs self.max_match = 0; # n workers and n jobs self.lx = [0]*31 # labels of X and Y parts self.ly = [0]*31 # labels of X and Y parts self.xy = [0]*31 # xy[x] - vertex that is matched with x, self.yx = [0]*31 # yx[y] - vertex that is matched with y self.S = [False]*31 # sets S and T in algorithm self.T = [False]*31 # sets S and T in algorithm self.slack = [0]*31 # as in the algorithm description self.slackx = [0]*31 # slackx[y] such a vertex, that self.prev_ious = [0]*31 # array for memorizing alternating p def init_labels(self): for i in range(len(self.lx)): self.lx[i] = 0 for i in range(len(self.ly)): self.ly[i] = 0 for x in range(self.n): for y in range(self.n): self.lx[x] = max(self.lx[x], self.cost[x][y]) def update_labels(self): x = 0 y = 0 delta = 99999999 # init delta as infinity for y in range(self.n): # calculate delta using slack if this.T[y] == False: delta = math.min(delta, self.slack[y]) for x in range(self.n): if self.S[x] == True: # update X labels self.lx[x] -= delta for y in range(self.n): if self.T[y] == True: self.ly[y] += delta # update Y labels for y in range(self.n): if self.T[y] == False: self.slack[y] -= delta # update slack array def add_to_tree(self, x, prev_iousx): # x - current vertex,prev_iousx - vertex from X before x in the alternating path, # so we add edges (prev_iousx, xy[x]), (xy[x], x) self.S[x] = True # add x to S self.prev_ious[x] = prev_iousx # we need this when augmenting for y in range(self.n): if self.lx[x] + self.ly[y] - self.cost[x][y] < self.slack[y]: self.slack[y] = self.lx[x] + self.ly[y] - self.cost[x][y] self.slackx[y] = x def augment(self): # main function of the algorithm if self.max_match == self.n: return # check whether matching is already perfect x= 0 y = 0 root = 0 # just counters and root vertex q= [0]*31 wr = 0 rd = 0 # q - queue for bfs, wr,rd - write and read # pos in queue for i in range(len(self.S)): self.S[i] = False # init set S for i in range(len(self.T)): self.T[i] = False # init set T for i in range(len(self.prev_ious)): self.prev_ious[i] = -1 # init set S for x in range(self.n): # finding root. if self.xy[x] == -1: q[wr] = root = x wr = wr + 1 self.prev_ious[x] = -2 self.S[x] = True break for y in range(self.n): #initializing slack array self.slack[y] = self.lx[root] + self.ly[y] - self.cost[root][y] self.slackx[y] = root # second part of augment() function while (True): # main cycle while (rd < wr): # building tree with bfs cycle x = q[rd] rd = rd + 1 # current vertex from X part for y in range(self.n): # iterate through all edges in equality graph if self.cost[x][y] == self.lx[x] + self.ly[y] and self.T[y] == False: if self.yx[y] == -1: break # an exposed vertex in Y found, so # augmenting path exists! this.T[y] = True # else just add y to T, q[wr] = self.yx[y] # add vertex yx[y], which is matched wr = wr + 1 # with y, to the queue self.add_to_tree(self.yx[y], x) # add edges (x,y) and (y,yx[y]) to the tree if y < self.n: break # augmenting path found! if y < self.n: break # augmenting path found! self.update_labels() #augmenting path not found, so improve labeling wr = 0 rd = 0 for y in range(self.n): if self.T[y] == False and self.slack[y] == 0: if self.yx[y] == -1: # exposed vertex in Y found - augmenting path exists! x = self.slackx[y] break else: self.T[y] = True # else just add y to T, if self.S[self.yx[y]] == False: q[wr] = self.yx[y] # add vertex yx[y], which is matched with wr = wr + 1 # y, to the queue self.add_to_tree(self.yx[y], self.slackx[y]); # and add edges (x,y) and (y, # //yx[y]) to the tree if y < self.n: break # augmenting path found! if y < self.n: # we found augmenting path! self.max_match = self.max_match + 1 # increment matching # in this cycle we inverse edges along augmenting path cx = x cy = y ty = 0 while(cx != -2): ty = self.xy[cx] self.yx[cy] = cx self.xy[cx] = cy cx = self.prev_ious[cx] cy = ty self.augment() #recall function, go to step 1 of the algorithm # end of augment() function def hungarian(self): ret = 0 # weight of the optimal matching self.max_match = 0; # number of vertices in current matching for i in range(len(self.xy)): self.xy[i] = -1 for i in range(len(self.yx)): self.yx[i] = -1 self.init_labels() # step 0 self.augment() # steps 1-3 for x in range(self.n): ret += self.cost[x][self.xy[x]] return ret def assignmentProblem(self, Arr, N): self.n = N for i in range(self.n): for j in range(self.n): self.cost[i][j] = -1*Arr[i*self.n + j] ans = -1 * self.hungarian() return ans - 2000n=3Arr = [1500,4000,4500,2000,6000,3500,2000,4000,2500] ob = Solution()print(ob.assignmentProblem( Arr,n))# The code is contributed by Nidhi goel. |
C#
using System;using System.Collections.Generic;using System.Collections;using System.Linq;// C# code implementationclass Solution { public int n; public int[][] cost = new int[31][]; //cost matrix public int max_match; //n workers and n jobs public int[] lx = new int[31]; public int[] ly = new int[31]; //labels of X and Y parts public int[] xy = new int[31]; //xy[x] - vertex that is matched with x, public int[] yx = new int[31]; //yx[y] - vertex that is matched with y public int[] S = new int[31]; public int[] T = new int[31]; //sets S and T in algorithm public int[] slack = new int[31]; //as in the algorithm description public int[] slackx = new int[31]; //slackx[y] such a vertex, that public int[] prev_ious = new int[31]; //array for memorizing alternating p public void init_labels() { for(int i = 0; i < lx.Length; i++) lx[i] = 0; for(int i = 0; i < ly.Length; i++) ly[i] = 0; for (int x = 0; x < n; x++) for (int y = 0; y < n; y++) lx[x] = Math.Max(lx[x], cost[x][y]); } public void update_labels() { int x, y; int delta = 99999999; //init delta as infinity for (y = 0; y < n; y++) //calculate delta using slack if (T[y] == 0) delta = Math.Min(delta, slack[y]); for (x = 0; x < n; x++) //update X labels if (S[x] == 1) lx[x] -= delta; for (y = 0; y < n; y++) //update Y labels if (T[y] == 1) ly[y] += delta; for (y = 0; y < n; y++) //update slack array if (T[y]==0) slack[y] -= delta; } public void add_to_tree(int x, int prev_iousx) //x - current vertex,prev_iousx - vertex from X before x in the alternating path, //so we add edges (prev_iousx, xy[x]), (xy[x], x) { S[x] = 1; //add x to S prev_ious[x] = prev_iousx; //we need this when augmenting for (int y = 0; y < n; y++) //update slacks, because we add new vertex to S if (lx[x] + ly[y] - cost[x][y] < slack[y]) { slack[y] = lx[x] + ly[y] - cost[x][y]; slackx[y] = x; } } public void augment() //main function of the algorithm { if (max_match == n) return; //check whether matching is already perfect int x, y; //just counters and root vertex int[] q = new int[n]; int wr = 0, rd = 0; //q - queue for bfs, wr,rd - write and read //pos in queue for(int i = 0; i < S.Length; i++) S[i] = 0; for(int i = 0; i < T.Length; i++) T[i] = 0; for(int i = 0; i < prev_ious.Length; i++) prev_ious[i] = -1; int root = -1; for (x = 0; x < n; x++) //finding root of the tree { if (xy[x] == -1) { q[wr] = root = x; wr = wr + 1; prev_ious[x] = -2; S[x] = 1; break; } } if (root == -1) { // All vertices are already matched return; } for (y = 0; y < n; y++) //initializing slack array { slack[y] = lx[root] + ly[y] - cost[root][y]; slackx[y] = root; } //second part of augment() function while (true) //main cycle { while (rd < wr) //building tree with bfs cycle { x = q[rd++]; //current vertex from X part for (y = 0; y < n; y++) //iterate through all edges in equality graph if (cost[x][y] == lx[x] + ly[y] && T[y]==0) { if (yx[y] == -1) break; //an exposed vertex in Y found, so //augmenting path exists! T[y] = 1; //else just add y to T, q[wr] = yx[y]; //add vertex yx[y], which is matched wr =wr +1; //with y, to the queue add_to_tree(yx[y], x); //add edges (x,y) and (y,yx[y]) to the tree } if (y < n) break; //augmenting path found! } if (y < n) break; //augmenting path found! update_labels(); //augmenting path not found, so improve labeling wr = rd = 0; for (y = 0; y < n; y++) //in this cycle we add edges that were added to the equality graph as a //result of improving the labeling, we add edge (slackx[y], y) to the tree if //and only if !T[y] && slack[y] == 0, also with this edge we add another one //(y, yx[y]) or augment the matching, if y was exposed if (T[y]==0 && slack[y] == 0) { if (yx[y] == -1) //exposed vertex in Y found - augmenting path exists! { x = slackx[y]; break; } else { T[y] = 1; //else just add y to T, if (S[yx[y]] ==0) { q[wr++] = yx[y]; //add vertex yx[y], which is matched with //y, to the queue add_to_tree(yx[y], slackx[y]); //and add edges (x,y) and (y, //yx[y]) to the tree } } } if (y < n) break; //augmenting path found! } if (y < n) //we found augmenting path! { max_match++; //increment matching //in this cycle we inverse edges along augmenting path for (int cx = x, cy = y, ty; cx != -2; cx = prev_ious[cx], cy = ty) { ty = xy[cx]; yx[cy] = cx; xy[cx] = cy; } augment(); //recall function, go to step 1 of the algorithm } }//end of augment() function public int hungarian() { int ret = 0; //weight of the optimal matching max_match = 0; //number of vertices in current matching xy = new int[n]; yx = new int[n]; for(int i = 0; i < xy.Length; i++) xy[i] = -1; for(int i = 0; i < yx.Length; i++) yx[i] = -1; init_labels(); //step 0 augment(); //steps 1-3 for (int x = 0; x < n; x++) //forming answer there ret += cost[x][xy[x]]; return ret; } public int assignmentProblem(int[] Arr, int N) { // initialising cost array. for(int i = 0; i < 31; i++){ cost[i] = new int[31]; } n = N; for(int i=0; i<n; i++) for(int j=0; j<n; j++) cost[i][j] = -1*Arr[i*n+j]; int ans = -1 * hungarian(); return ans; } static void Main() { int n=3; int[] Arr={1500,4000,4500,2000,6000,3500,2000,4000,2500}; /*1500 4000 4500 2000 6000 3500 2000 4000 2500*/ Solution ob = new Solution(); Console.WriteLine(ob.assignmentProblem(Arr,n)); }}// The code is contributed by Nidhi goel. |
Javascript
// Javascript implementation of the above approach.class Solution { constructor(){ // cost matrix this.cost = new Array(31); for(let i = 0; i < 31; i++){ this.cost[i] = new Array(31).fill(0); } this.n = 0; //n workers and n jobs this.max_match = 0; //n workers and n jobs this.lx = new Array(31).fill(0); //labels of X and Y parts this.ly = new Array(31).fill(0); //labels of X and Y parts this.xy = new Array(31).fill(0); //xy[x] - vertex that is matched with x, this.yx = new Array(31).fill(0); //yx[y] - vertex that is matched with y this.S = new Array(31).fill(false); //sets S and T in algorithm this.T = new Array(31).fill(false); //sets S and T in algorithm this.slack = new Array(31).fill(0); //as in the algorithm description this.slackx = new Array(31).fill(0); //slackx[y] such a vertex, that this.prev_ious = new Array(31).fill(0); //array for memorizing alternating p } init_labels() { for(let i = 0; i < this.lx.length; i++) this.lx[i] = 0; for(let i = 0; i < this.ly.length; i++) this.ly[i] = 0; for (let x = 0; x < this.n; x++) for (let y = 0; y < this.n; y++) this.lx[x] = Math.max(this.lx[x], this.cost[x][y]); } update_labels() { let x = 0; let y = 0; let delta = 99999999; //init delta as infinity for (y = 0; y < this.n; y++) //calculate delta using slack if (!this.T[y]) delta = Math.min(delta, this.slack[y]); for (x = 0; x < this.n; x++) //update X labels if (this.S[x]) this.lx[x] -= delta; for (y = 0; y < this.n; y++) //update Y labels if (this.T[y]) this.ly[y] += delta; for (y = 0; y < this.n; y++) //update slack array if (!this.T[y]) this.slack[y] -= delta; } add_to_tree(x, prev_iousx) //x - current vertex,prev_iousx - vertex from X before x in the alternating path, //so we add edges (prev_iousx, xy[x]), (xy[x], x) { this.S[x] = true; //add x to S this.prev_ious[x] = prev_iousx; //we need this when augmenting for (let y = 0; y < this.n; y++) //update slacks, because we add new vertex to S if (this.lx[x] + this.ly[y] - this.cost[x][y] < this.slack[y]) { this.slack[y] = this.lx[x] + this.ly[y] - this.cost[x][y]; this.slackx[y] = x; } } augment() //main function of the algorithm { if (this.max_match == this.n) return; //check whether matching is already perfect let x= 0; let y = 0; let root = 0; //just counters and root vertex let q= new Array(31); let wr = 0; let rd = 0; //q - queue for bfs, wr,rd - write and read //pos in queue for(let i = 0; i < this.S.length; i++) this.S[i] = false; //init set S for(let i = 0; i < this.T.length; i++) this.T[i] = false; //init set T for(let i = 0; i < this.prev_ious.length; i++) this.prev_ious[i] = -1; //init set prev_ious - for the alternating tree for (x = 0; x < this.n; x++) //finding root of the tree { if (this.xy[x] == -1) { q[wr++] = root = x; this.prev_ious[x] = -2; this.S[x] = true; break; } } for (y = 0; y < this.n; y++) //initializing slack array { this.slack[y] = this.lx[root] + this.ly[y] - this.cost[root][y]; this.slackx[y] = root; } //second part of augment() function while (true) //main cycle { while (rd < wr) //building tree with bfs cycle { x = q[rd++]; //current vertex from X part for (y = 0; y < this.n; y++) //iterate through all edges in equality graph if (this.cost[x][y] == this.lx[x] + this.ly[y] && !this.T[y]) { if (this.yx[y] == -1) break; //an exposed vertex in Y found, so //augmenting path exists! this.T[y] = true; //else just add y to T, q[wr++] = this.yx[y]; //add vertex yx[y], which is matched //with y, to the queue this.add_to_tree(this.yx[y], x); //add edges (x,y) and (y,yx[y]) to the tree } if (y < this.n) break; //augmenting path found! } if (y < this.n) break; //augmenting path found! this.update_labels(); //augmenting path not found, so improve labeling wr = rd = 0; for (y = 0; y < this.n; y++) // in this cycle we add edges that were added to the equality graph as a // result of improving the labeling, we add edge (slackx[y], y) to the tree if // and only if !T[y] && slack[y] == 0, also with this edge we add another one // (y, yx[y]) or augment the matching, if y was exposed if (!this.T[y] && this.slack[y] == 0) { if (this.yx[y] == -1) //exposed vertex in Y found - augmenting path exists! { x = this.slackx[y]; break; } else { this.T[y] = true; //else just add y to T, if (!this.S[this.yx[y]]) { q[wr++] = this.yx[y]; //add vertex yx[y], which is matched with //y, to the queue this.add_to_tree(this.yx[y], this.slackx[y]); //and add edges (x,y) and (y, //yx[y]) to the tree } } } if (y < this.n) break; //augmenting path found! } if (y < this.n) //we found augmenting path! { this.max_match++; //increment matching // in this cycle we inverse edges along augmenting path for (let cx = x, cy = y, ty; cx != -2; cx = this.prev_ious[cx], cy = ty) { ty = this.xy[cx]; this.yx[cy] = cx; this.xy[cx] = cy; } this.augment(); //recall function, go to step 1 of the algorithm } }// end of augment() function hungarian() { let ret = 0; //weight of the optimal matching this.max_match = 0; //number of vertices in current matching for(let i= 0; i < this.xy.length; i++) this.xy[i] = -1; for(let i = 0; i < this.yx.length; i++) this.yx[i] = -1; this.init_labels(); //step 0 this.augment(); //steps 1-3 for (let x = 0; x < this.n; x++) //forming answer there ret += this.cost[x][this.xy[x]]; return ret; } assignmentProblem(Arr, N) { this.n = N; for(let i=0; i<this.n; i++) for(let j=0; j<this.n; j++) this.cost[i][j] = -1*Arr[i*this.n+j]; let ans = -1 * this.hungarian(); return ans; }};let n=3;let Arr = [1500,4000,4500,2000,6000,3500,2000,4000,2500]; /*1500 4000 4500 2000 6000 3500 2000 4000 2500*/let ob = new Solution();console.log(ob.assignmentProblem(Arr,n));// The code is contributed by Arushi Jindal. |
Output
8500
Time complexity : O(n^3), where n is the number of workers and jobs. This is because the algorithm implements the Hungarian algorithm, which is known to have a time complexity of O(n^3).
Space complexity : O(n^2), where n is the number of workers and jobs. This is because the algorithm uses a 2D cost matrix of size n x n to store the costs of assigning each worker to a job, and additional arrays of size n to store the labels, matches, and auxiliary information needed for the algorithm.
In the next post, we will be discussing implementation of the above algorithm. The implementation requires more steps as we need to find minimum number of lines to cover all 0’s using a program. References: http://www.math.harvard.edu/archive/20_spring_05/handouts/assignment_overheads.pdf https://www.youtube.com/watch?v=dQDZNHwuuOY This article is contributed by Yash Varyani. Please write comments if you find anything incorrect, or you want to share more information about the topic discussed above.
Please Login to comment...