# Simulated Annealing

In relation to another question I had, I have been researching Simulated Annealing.

The general example used with this algorithm is the traveling salesman example. I have been testing the code described here : https://jamesmccaffrey.wordpress.com/2021/12/01/the-traveling-salesman-problem-using-simulated-annealing-in-csharp/

I have noticed something rather annoying though.

If I set the number of cites to 60, with a maxIteration count of 200k, an alpha of 0.95 (rate at which temperature decreases) and a seed of 2, I get a good result.

It finds the optional solution in 143k iterations

If I then decrease the number of cities to 59 (with the same settings), I get a (very) poor result.

It maxes out the iterations with a solution more than twice the distance of the optimal solution

Simply changing the seed gives widely varying results.

I appreciate that this is to be expected due to the nature of randomness. But I am wondering if this is an implementation issue making things worse or just the nature of the algorithm?

Here is the C# code :

``````namespace TSP_Annealing
{
class TSP_Annealing_Program
{
static void Main(string[] args)
{
int nCities = 60;
Random rnd = new Random(2);
int maxIter = 200000;
double startTemperature = 10000.0;
double alpha = 0.95;

(int[] soln, int iteration) = Solve(nCities, rnd, maxIter, startTemperature, alpha);
Console.WriteLine(\$"Finished solve({iteration}) ");

Console.WriteLine("\nBest solution found: ");
ShowArray(soln);
double dist = TotalDist(soln);
Console.WriteLine("\nTotal distance = " +  dist.ToString("F1"));

}

static double TotalDist(int[] route)
{
double d = 0.0;  // total distance between cities
int n = route.Length;
for (int i = 0; i < n - 1; ++i)
{
if (route[i] < route[i + 1])
d += (route[i + 1] - route[i]) * 1.0;
else
d += (route[i] - route[i + 1]) * 1.5;
}
return d;
}

static double Error(int[] route)
{
int n = route.Length;
double d = TotalDist(route);
double minDist = n - 1;
return d - minDist;
}

static int[] Adjacent(int[] route, Random rnd)
{
int n = route.Length;
int[] result = (int[])route.Clone();  // shallow is OK
int i = rnd.Next(0, n); int j = rnd.Next(0, n);
int tmp = result[i];
result[i] = result[j]; result[j] = tmp;
return result;
}

static void Shuffle(int[] route, Random rnd)
{
// Fisher-Yates algorithm
int n = route.Length;
for (int i = 0; i < n; ++i)
{
int rIndx = rnd.Next(i, n);
int tmp = route[rIndx];
route[rIndx] = route[i];
route[i] = tmp;
}
}

static void ShowArray(int[] arr)
{
int n = arr.Length;
Console.Write("[ ");
for (int i = 0; i < n; ++i)
Console.WriteLine(" ]");
}

static (int[], int totalIterations) Solve(int nCities, Random rnd, int maxIter, double startTemperature, double alpha)
{
double currTemperature = startTemperature;
int[] soln = new int[nCities];
for (int i = 0; i < nCities; ++i) { soln[i] = i; }
Shuffle(soln, rnd);
Console.WriteLine("Initial guess: ");
ShowArray(soln);

double err = Error(soln);
int iteration = 0;
while (iteration < maxIter && err > 0.0)
{
if (adjErr < err)  // better route
{
}
else
{
double acceptProb =
double p = rnd.NextDouble(); // corrected
if (p < acceptProb)  // accept anyway
{
}
}

if (currTemperature < 0.00001)
currTemperature = 0.00001;
else
currTemperature *= alpha;
++iteration;
}

return (soln, iteration);
}

}
}
``````

Edit : I don't appear to be having much luck with this. As suggested, I tried writing the example again, but this time in 2D. To get the next route, I followed the suggestion in this tutorial https://youtu.be/AEeYp5VtI08?t=1174 which picks a random segment and reverses it.

The example seems to work when I set the random seed to 2 and finds the shortest distance of ~1695, but any other seed results in it getting stuck with distances much greater.

``````namespace TSP_Annealing
{
public class Map
{
public const int Size = 1000;
public List<Point> Cities = new();

public Map(int cities)
{
var r = new Random(0);
for (int i = 0; i < cities; i++)
{
int x = r.Next(0, Size);
int y = r.Next(0, Size);

}
}
}

public class TSM
{
private Map _map = new Map(8);
private const int _maxIterations = 20000;
private const double _startTemperature = 10000;
private const double _alpha = 0.95;
private Random _rnd = new Random(2);

public void Solve()
{
double currTemp = _startTemperature;
var shortestRoute = Shuffle(Enumerable.Range(0, _map.Cities.Count).ToArray());
double shortestDistance = ComputeDistance(shortestRoute);

int iteration = 0;

while (iteration < _maxIterations)
{
int[] nextRoute = NextRoute(shortestRoute);
double nextDistance = ComputeDistance(nextRoute);

if (nextDistance < shortestDistance)
{
shortestRoute = nextRoute;
shortestDistance = nextDistance;
}
else
{
double acceptProb = Math.Exp(-(nextDistance - shortestDistance) / currTemp);
double p = _rnd.NextDouble();
if (p < acceptProb)
{
shortestRoute = nextRoute;
shortestDistance = nextDistance;
}
}

if (currTemp < 0.00001)
currTemp = 0.00001;
else
currTemp *= _alpha;

++iteration;
}

Debug.WriteLine(\$"Shortest : {shortestDistance}, Iterations: {iteration}");
}

private double ComputeDistance(int[] route)
{
double total = 0.0;

for (int i = 0; i < route.Length - 1; ++i)
{
Point a = _map.Cities[route[i]];
Point b = _map.Cities[route[i + 1]];
double x = b.X - a.X;
double y = b.Y - a.Y;
double length = Math.Sqrt(x * x + y * y);
total += length;
}

}

private int[] NextRoute(int[] route)
{
// pick two random points to define a segment
int p1 = _rnd.Next(0, route.Length);
int p2 = _rnd.Next(p1, route.Length);

while ((p2 - p1) <= 1)
{
p1 = _rnd.Next(0, route.Length);
p2 = _rnd.Next(p1, route.Length);
}

// reverse the segment
var copy = (int[])route.Clone();
for (int i = 0; i < route.Length; ++i)
{
int n = p1 + i;
int m = p2 - i;

if (n >= m) break;

int temp = copy[n];
copy[n] = copy[m];
copy[m] = temp;
}

return copy;
}

private int[] Shuffle(int[] route)
{
int n = route.Length;
for (int i = 0; i < n; ++i)
{
int rIndx = _rnd.Next(i, n);
int tmp = route[rIndx];
route[rIndx] = route[i];
route[i] = tmp;
}

return route;
}
}

class TSP_Annealing_Program
{
static void Main(string[] args)
{
var tsm = new TSM();
tsm.Solve();
}
}
}
``````
• Let's say you have 5 local minimums and one global minimum. What's to stop you landing in a good place to begin with, but then as the temperature is still high, you jump to a much worse different local minimum and then get stuck there as the temperature cools? Aug 28 at 21:30

I think the main issue with your first TSP approach is that this test problem is a one dimensional variant of TSP with asymmetric costs (costs for A to B are not the same as for B to A).

This is a very stiff problem - local optima tend to consist of a few monotonic sequence strains, and the shown `Adjacent` function for making local modifications is not suited well to break up these strains, or turn their directions, which would be necessary to get out of a local optimum.

This kind of setup makes it hard for almost any heuristic algorithm using random changes to get out of a local optimum once it got stuck there, even when it allows the target function to increase its values with a certain probability. As a solution, one can try to find a better `Adjacent` function, which is way easier when we look at TSPs with symmetric costs first.

The classic TSP model problem on a 2D map uses typically euclidean distances, which are symmetric. For this setup, it is straightforward to implement an `Adjacent` function which cuts out a partial route in the middle, inverts the direction of that route and reinserts the inverted sequence (so it actually swaps two cities and reverts the route between them, which does not change the costs in the middle). I guess that will also work for a 1D TSP with symmetric costs.

In short, this kind of "model TSP" together with the shown `Adjacent` function is not optimal for demonstrating the capabilities of evolutionary optimization algorithms like SA. You need to find a better `Adjacent` function, or a different model problem.

Your second TSP model problem may be better suited, but it shows some other potential obstacles:

• when starting with a too high temperature, SA will just shuffle the routes randomly for a long time, so even a higher number of iterations won't help much
• when decreasing the temperature too quickly, SA will too early get stuck into a local minimum.
• when decreasing the temperature too slowly, one has to increase the number of iterations heavily to reach the sufficiently low temperatures where the total distance settles

One has to find the sweet spot where the algorithm finds local improvements most times, but jumps into a different area of the search space with a certain probability from time to time. That is is the temperature area where the algorithm should reside with most of its iterations.

Hence, one has to tailor the starting temperature, cooling scheme, number of iterations to the specific problem.

• bah, not having much luck. I have updated the question with a new 2d version I wrote, but its getting stuck most of the time. Aug 28 at 17:58
• Am I right in understanding that no matter what seed I set it to, I should be able to reach the shortest distance? it might just mean I have to increase the iterations? I seem to be getting stuck in a local minimum Aug 28 at 18:18
• @wforl: em, no. SA is not an algorithm for always finding the global optimum. But for many problems, it can find pretty "good" solutions. For your 2nd implementation, increasing number of iterations and decreasing the temperature more slowly might help. And I can confirm all of what njuffa wrote in their comments. Aug 29 at 6:38
• If I reduce the temperature more slowly it improves one case and makes another case worse. It seems like in all examples of the algorithm, they never keep track of the lowest value. For example if you jump from one local minimum to a worse one, you still ending up returning the the worse one at the end if you get stuck in it. Maybe it should keep track of a global best? Such that if it dies jump around, and gets worse, it can still return an earlier value? Aug 29 at 6:54
• I'll experiment with adding more cycles so it can reset, but I simply don't understand why everyone else seems to use SA with TSM with great luck and yet I'm getting rubbish results with only 8 cities. Aug 29 at 6:58