44.Prim算法

public static void main(String[] args) {
      //测试看看图是否创建ok
      char[] data = new char[]{'A','B','C','D','E','F','G'};
      int verxs = data.length;
      //邻接矩阵的关系使用二维数组表示,10000这个大数,表示两个点不联通
      int [][]weight=new int[][]{
            {10000,5,7,10000,10000,10000,2},
            {5,10000,10000,9,10000,10000,3},
            {7,10000,10000,10000,8,10000,10000},
            {10000,9,10000,10000,10000,4,10000},
            {10000,10000,8,10000,10000,5,4},
            {10000,10000,10000,4,5,10000,6},
            {2,3,10000,10000,4,6,10000},};
            
        //创建MGraph对象
        MGraph graph = new MGraph(verxs);
        //创建一个MinTree对象
        MinTree minTree = new MinTree();
        minTree.createGraph(graph, verxs, data, weight);
        //输出
        minTree.showGraph(graph);
        //测试普利姆算法
        minTree.prim(graph, 1);// 
   }

}

//创建最小生成树->村庄的图
class MinTree {
   //创建图的邻接矩阵
   /**
    * 
    * @param graph 图对象
    * @param verxs 图对应的顶点个数
    * @param data 图的各个顶点的值
    * @param weight 图的邻接矩阵
    */
   public void createGraph(MGraph graph, int verxs, char data[], int[][] weight) {
      int i, j;
      for(i = 0; i < verxs; i++) {//顶点
         graph.data[i] = data[i];
         for(j = 0; j < verxs; j++) {
            graph.weight[i][j] = weight[i][j];
         }
      }
   }
   
   //显示图的邻接矩阵
   public void showGraph(MGraph graph) {
      for(int[] link: graph.weight) {
         System.out.println(Arrays.toString(link));
      }
   }
   
   //编写prim算法,得到最小生成树
   /**
    * 
    * @param graph 图
    * @param v 表示从图的第几个顶点开始生成'A'->0 'B'->1...
    */
   public void prim(MGraph graph, int v) {
      //visited[] 标记结点(顶点)是否被访问过
      int visited[] = new int[graph.verxs];
      //visited[] 默认元素的值都是0, 表示没有访问过
//    for(int i =0; i <graph.verxs; i++) {
//       visited[i] = 0;
//    }
      
      //把当前这个结点标记为已访问
      visited[v] = 1;
      //h1 和 h2 记录两个顶点的下标
      int h1 = -1;
      int h2 = -1;
      int minWeight = 10000; //将 minWeight 初始成一个大数,后面在遍历过程中,会被替换
      for(int k = 1; k < graph.verxs; k++) {//因为有 graph.verxs顶点,普利姆算法结束后,有 graph.verxs-1边
         
         //这个是确定每一次生成的子图 ,和哪个结点的距离最近
         for(int i = 0; i < graph.verxs; i++) {// i结点表示被访问过的结点
            for(int j = 0; j< graph.verxs;j++) {//j结点表示还没有访问过的结点
               if(visited[i] == 1 && visited[j] == 0 && graph.weight[i][j] < minWeight) {
                  //替换minWeight(寻找已经访问过的结点和未访问过的结点间的权值最小的边)
                  minWeight = graph.weight[i][j];
                  h1 = i;
                  h2 = j;
               }
            }
         }
         //找到一条边是最小
         System.out.println("边<" + graph.data[h1] + "," + graph.data[h2] + "> 权值:" + minWeight);
         //将当前这个结点标记为已经访问
         visited[h2] = 1;
         //minWeight 重新设置为最大值 10000
         minWeight = 10000;
      }
      
   }
}

class MGraph {
   int verxs; //表示图的节点个数
   char[] data;//存放结点数据
   int[][] weight; //存放边,就是我们的邻接矩阵
   
   public MGraph(int verxs) {
      this.verxs = verxs;
      data = new char[verxs];
      weight = new int[verxs][verxs];
   }
}
上一篇:ModelAndView


下一篇:35.图