How to improve Dijkstra algorithm when querying n times? - algorithm

Question

I'm currently working on a problem at Codechef. You can find the problem statement here:
Delivery Boy
In short, the problem is asking to query n times the shortest path from a start to an end. My solution is to use Dijsktra with priority_queue plus caching the result into a hash_map in case we already had a start. Unfortunately, I got time limit exceed many times and I couldn't find a better way to make it faster. I wonder am I in the right track? or there is a better algorithm to this problem?
By the way, since the contest is still going, please don't post any solution. A hint is more than enough to me. Thanks.
Here is my attempt:
#ifdef __GNUC__
#include <ext/hash_map>
#else
#include <hash_map>
#endif
#include <iostream>
#include <iomanip>
#include <vector>
#include <string>
#include <algorithm>
#include <map>
#include <set>
#include <utility>
#include <stack>
#include <deque>
#include <queue>
#include <fstream>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <cmath>
#include <cassert>
using namespace std;
#ifdef __GNUC__
namespace std {
using namespace __gnu_cxx;
}
#endif
const int MAX_VERTICES = 250;
const int INFINIY = (1 << 28);
int weight[MAX_VERTICES + 1][MAX_VERTICES + 1];
bool visited_start[MAX_VERTICES + 1] = { 0 };
struct vertex {
int node;
int cost;
vertex(int node = 0, int cost = 0)
: node(node), cost(cost) {
}
bool operator <(const vertex& rhs) const {
return cost < rhs.cost;
}
bool operator >(const vertex& rhs) const {
return cost > rhs.cost;
}
};
hash_map<int, vector<vertex> > cache;
typedef priority_queue<vertex, vector<vertex>, greater<vertex> > min_pq;
vector<vertex> dijkstra_compute_path(int start, int n) {
min_pq pq;
vector<vertex> path;
vector<int> visited(n, 0);
int min_cost = 0;
int better_cost;
vertex u;
for (int i = 0; i < n; ++i) {
path.push_back(vertex(i, INFINIY));
}
path[start].cost = 0;
pq.push(vertex(start, path[start].cost));
while (!pq.empty()) {
// extract min cost
u = pq.top();
pq.pop();
// mark it as visited
visited[u.node] = 1;
// for each vertex v that is adjacent to u
for (int v = 0; v < n; ++v) {
// if it's not visited, visit it
if (visited[v] == 0) {
better_cost = path[u.node].cost + weight[u.node][v];
// update cost
if (path[v].cost > better_cost) {
path[v].cost = better_cost;
pq.push(vertex(v, path[v].cost));
}
}
}
}
return path;
}
void check_in_cache(vector<vertex>& path, int start, int no_street) {
if (visited_start[start] == 0) {
path = dijkstra_compute_path(start, no_street);
cache.insert(make_pair(start, path));
visited_start[start] = 1;
}
else {
path = cache[start];
}
}
void display_cost(int stop_at_gas_cost, int direct_cost) {
printf("%d ", stop_at_gas_cost);
if (stop_at_gas_cost > direct_cost) {
printf("%d\n", stop_at_gas_cost - direct_cost);
}
else {
printf("0\n");
}
}
void handle_case_one() {
int no_scenario;
int dummy;
int s, g, d;
scanf("%d", &dummy);
scanf("%d", &no_scenario);
for (int i = 0; i < no_scenario; ++i) {
scanf("%d %d %d", &s, &g, &d);
printf("0 0\n");
}
}
void inout_delivery_boy() {
int no_street;
int no_scenario;
int restaurant;
int gas_station;
int destination;
int stop_at_gas_cost;
int direct_cost;
vector<vertex> direct;
vector<vertex> indirect;
vector<vertex> d;
int c;
scanf("%d", &no_street);
if (no_street == 1) {
handle_case_one();
return;
}
for (int x = 0; x < no_street; ++x) {
for (int y = 0; y < no_street; ++y) {
scanf("%d", &c);
weight[x][y] = c;
}
}
for (int i = 0; i < no_street; ++i) {
d.push_back(vertex(i, INFINIY));
}
scanf("%d", &no_scenario);
for (int i = 0; i < no_scenario; ++i) {
scanf("%d %d %d", &restaurant, &gas_station, &destination);
// check in cache
check_in_cache(direct, restaurant, no_street);
check_in_cache(indirect, gas_station, no_street);
// calculate the cost
stop_at_gas_cost = direct[gas_station].cost + indirect[destination].cost;
direct_cost = direct[destination].cost;
// output
display_cost(stop_at_gas_cost, direct_cost);
}
}
void dijkstra_test(istream& in) {
int start;
int no_street;
int temp[4] = { 0 };
vector<vertex> path;
in >> no_street;
for (int x = 0; x < no_street; ++x) {
for (int y = 0; y < no_street; ++y) {
in >> weight[x][y];
}
}
// arrange
start = 0;
temp[0] = 0;
temp[1] = 2;
temp[2] = 1;
temp[3] = 3;
// act
path = dijkstra_compute_path(start, no_street);
// assert
for (int i = 0; i < no_street; ++i) {
assert(path[i].cost == temp[i]);
}
// arrange
start = 1;
temp[0] = 1;
temp[1] = 0;
temp[2] = 2;
temp[3] = 4;
// act
path = dijkstra_compute_path(start, no_street);
// assert
for (int i = 0; i < no_street; ++i) {
assert(path[i].cost == temp[i]);
}
// arrange
start = 2;
temp[0] = 2;
temp[1] = 1;
temp[2] = 0;
temp[3] = 3;
// act
path = dijkstra_compute_path(start, no_street);
// assert
for (int i = 0; i < no_street; ++i) {
assert(path[i].cost == temp[i]);
}
// arrange
start = 3;
temp[0] = 1;
temp[1] = 1;
temp[2] = 1;
temp[3] = 0;
// act
path = dijkstra_compute_path(start, no_street);
// assert
for (int i = 0; i < no_street; ++i) {
assert(path[i].cost == temp[i]);
}
}
int main() {
// ifstream inf("test_data.txt");
// dijkstra_test(inf);
inout_delivery_boy();
return 0;
}

Answer 1

please notice N is small in the problem. have you tried Floyd shortest path algorithm to pre-calculate shortest path between each two nodes ? it will cost O(N^3) time, which is 250^3=15625000 in the problem, should be easy to be finished running in 1 second. Then you can answer each query in O(1).
the introduction of Floyd :
http://en.wikipedia.org/wiki/Floyd%E2%80%93Warshall_algorithm
ps: i think cached dijstra costs a maximum running time of O(N^3) for overall test case as well . but the way you implement the cache will spend more unnecessary time on memory copying, which may lead to a TLE. Just a guess.

Answer 2

Indeed Floyd-Warshall's Algorithm is better than Dijkstra's in this case, the complexity for Dijkstra is O(m*n^2) and in this problem M is much much higher than N so the O(n^3) time complexity of Floyd-Warshall is better.