Rewrite TSP solver for improved path optimization and clarity
- Completely re-implemented the TSP algorithm in C++ for better path quality - Added detailed comments and documentation to clarify each step - Improved nearest neighbor, 2-opt, and relocation logic - Enhanced handling of start/end point constraints - Updated PathUtils.py docstring to accurately describe start point behavior
This commit is contained in:
Billy Huddleston
@@ -28,7 +28,15 @@
|
||||
|
||||
namespace
|
||||
{
|
||||
// Euclidean distance between two points
|
||||
/**
|
||||
* @brief Calculate Euclidean distance between two points
|
||||
*
|
||||
* Used for 2-opt and relocation steps where actual distance matters for path length optimization.
|
||||
*
|
||||
* @param a First point
|
||||
* @param b Second point
|
||||
* @return Actual distance: sqrt(dx² + dy²)
|
||||
*/
|
||||
double dist(const TSPPoint& a, const TSPPoint& b)
|
||||
{
|
||||
double dx = a.x - b.x;
|
||||
@@ -36,22 +44,251 @@ double dist(const TSPPoint& a, const TSPPoint& b)
|
||||
return std::sqrt(dx * dx + dy * dy);
|
||||
}
|
||||
|
||||
// Calculate total path length
|
||||
double pathLength(const std::vector<TSPPoint>& points, const std::vector<int>& path)
|
||||
/**
|
||||
* @brief Calculate squared distance between two points (no sqrt)
|
||||
*
|
||||
* Used for nearest neighbor selection for performance (avoids expensive sqrt operation).
|
||||
* Since we only need to compare distances, squared distance preserves ordering:
|
||||
* if dist(A,B) < dist(A,C), then distSquared(A,B) < distSquared(A,C)
|
||||
*
|
||||
* @param a First point
|
||||
* @param b Second point
|
||||
* @return Squared distance: dx² + dy²
|
||||
*/
|
||||
double distSquared(const TSPPoint& a, const TSPPoint& b)
|
||||
{
|
||||
double total = 0.0;
|
||||
size_t n = path.size();
|
||||
for (size_t i = 0; i < n - 1; ++i) {
|
||||
total += dist(points[path[i]], points[path[i + 1]]);
|
||||
}
|
||||
// Optionally, close the loop: total += dist(points[path[n-1]], points[path[0]]);
|
||||
return total;
|
||||
double dx = a.x - b.x;
|
||||
double dy = a.y - b.y;
|
||||
return dx * dx + dy * dy;
|
||||
}
|
||||
|
||||
// 2-Opt swap
|
||||
void twoOptSwap(std::vector<int>& path, size_t i, size_t k)
|
||||
/**
|
||||
* @brief Core TSP solver implementation using nearest neighbor + iterative improvement
|
||||
*
|
||||
* Algorithm steps:
|
||||
* 1. Add temporary start/end points if specified
|
||||
* 2. Build initial route using nearest neighbor heuristic
|
||||
* 3. Optimize route with 2-opt and relocation moves
|
||||
* 4. Remove temporary points and map back to original indices
|
||||
*
|
||||
* @param points Input points to visit
|
||||
* @param startPoint Optional starting location constraint
|
||||
* @param endPoint Optional ending location constraint
|
||||
* @return Vector of indices representing optimized visit order
|
||||
*/
|
||||
std::vector<int> solve_impl(const std::vector<TSPPoint>& points,
|
||||
const TSPPoint* startPoint,
|
||||
const TSPPoint* endPoint)
|
||||
{
|
||||
std::reverse(path.begin() + static_cast<long>(i), path.begin() + static_cast<long>(k) + 1);
|
||||
// ========================================================================
|
||||
// STEP 1: Prepare point set with temporary start/end markers
|
||||
// ========================================================================
|
||||
// We insert temporary points to enforce start/end constraints.
|
||||
// These will be removed after optimization and won't appear in final result.
|
||||
std::vector<TSPPoint> pts = points;
|
||||
int tempStartIdx = -1, tempEndIdx = -1;
|
||||
|
||||
if (startPoint) {
|
||||
// Insert user-specified start point at beginning
|
||||
pts.insert(pts.begin(), TSPPoint(startPoint->x, startPoint->y));
|
||||
tempStartIdx = 0;
|
||||
}
|
||||
else if (!pts.empty()) {
|
||||
// No start specified: duplicate first point as anchor
|
||||
pts.insert(pts.begin(), TSPPoint(pts[0].x, pts[0].y));
|
||||
tempStartIdx = 0;
|
||||
}
|
||||
|
||||
if (endPoint) {
|
||||
// Add user-specified end point at the end
|
||||
pts.push_back(TSPPoint(endPoint->x, endPoint->y));
|
||||
tempEndIdx = static_cast<int>(pts.size()) - 1;
|
||||
}
|
||||
|
||||
// ========================================================================
|
||||
// STEP 2: Build initial route using Nearest Neighbor algorithm
|
||||
// ========================================================================
|
||||
// Greedy approach: always visit the closest unvisited point next.
|
||||
// This gives a decent initial solution quickly (O(n²) complexity).
|
||||
//
|
||||
// Tie-breaking rule:
|
||||
// - If distances are within ±0.1, prefer point with y-value closer to start
|
||||
// - This provides deterministic results when points are nearly equidistant
|
||||
std::vector<int> route;
|
||||
std::vector<bool> visited(pts.size(), false);
|
||||
route.push_back(0); // Start from temp start point (index 0)
|
||||
visited[0] = true;
|
||||
|
||||
for (size_t step = 1; step < pts.size(); ++step) {
|
||||
double minDist = std::numeric_limits<double>::max();
|
||||
int next = -1;
|
||||
double nextYDiff = std::numeric_limits<double>::max();
|
||||
|
||||
// Find nearest unvisited neighbor
|
||||
for (size_t i = 0; i < pts.size(); ++i) {
|
||||
if (!visited[i]) {
|
||||
// Use squared distance for speed (no sqrt needed for comparison)
|
||||
double d = distSquared(pts[route.back()], pts[i]);
|
||||
double yDiff = std::abs(pts[route.front()].y - pts[i].y);
|
||||
|
||||
// Tie-breaking logic:
|
||||
if (d > minDist + 0.1) {
|
||||
continue; // Clearly farther, skip
|
||||
}
|
||||
else if (d < minDist - 0.1) {
|
||||
// Clearly closer, use it
|
||||
minDist = d;
|
||||
next = static_cast<int>(i);
|
||||
nextYDiff = yDiff;
|
||||
}
|
||||
else if (yDiff < nextYDiff) {
|
||||
// Tie: prefer point closer to start in Y-axis
|
||||
minDist = d;
|
||||
next = static_cast<int>(i);
|
||||
nextYDiff = yDiff;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (next == -1) {
|
||||
break; // No more unvisited points
|
||||
}
|
||||
route.push_back(next);
|
||||
visited[next] = true;
|
||||
}
|
||||
|
||||
// Ensure temporary end point is at the end of route
|
||||
if (tempEndIdx != -1 && route.back() != tempEndIdx) {
|
||||
auto it = std::find(route.begin(), route.end(), tempEndIdx);
|
||||
if (it != route.end()) {
|
||||
route.erase(it);
|
||||
}
|
||||
route.push_back(tempEndIdx);
|
||||
}
|
||||
|
||||
// ========================================================================
|
||||
// STEP 3: Iterative improvement using 2-Opt and Relocation
|
||||
// ========================================================================
|
||||
// Repeatedly apply local optimizations until no improvement is possible.
|
||||
// This typically converges quickly (a few iterations) to a near-optimal solution.
|
||||
//
|
||||
// Two optimization techniques:
|
||||
// 1. 2-Opt: Reverse segments of the route to eliminate crossing paths
|
||||
// 2. Relocation: Move individual points to better positions in the route
|
||||
bool improvementFound = true;
|
||||
while (improvementFound) {
|
||||
improvementFound = false;
|
||||
|
||||
// --- 2-Opt Optimization ---
|
||||
// Try reversing every possible segment of the route.
|
||||
// If reversing segment [i+1...j-1] reduces total distance, keep it.
|
||||
//
|
||||
// Example: Route A-B-C-D-E becomes A-D-C-B-E if reversing B-C-D is better
|
||||
bool reorderFound = true;
|
||||
while (reorderFound) {
|
||||
reorderFound = false;
|
||||
for (size_t i = 0; i + 3 < route.size(); ++i) {
|
||||
for (size_t j = i + 3; j < route.size(); ++j) {
|
||||
// Current edges: i→(i+1) and (j-1)→j
|
||||
double curLen = dist(pts[route[i]], pts[route[i + 1]])
|
||||
+ dist(pts[route[j - 1]], pts[route[j]]);
|
||||
|
||||
// New edges after reversal: (i+1)→j and i→(j-1)
|
||||
// Add epsilon to prevent cycles from floating point errors
|
||||
double newLen = dist(pts[route[i + 1]], pts[route[j]])
|
||||
+ dist(pts[route[i]], pts[route[j - 1]]) + 1e-5;
|
||||
|
||||
if (newLen < curLen) {
|
||||
// Reverse the segment between i+1 and j (exclusive)
|
||||
std::reverse(route.begin() + i + 1, route.begin() + j);
|
||||
reorderFound = true;
|
||||
improvementFound = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// --- Relocation Optimization ---
|
||||
// Try moving each point to a different position in the route.
|
||||
// If moving point i to position j improves the route, do it.
|
||||
bool relocateFound = true;
|
||||
while (relocateFound) {
|
||||
relocateFound = false;
|
||||
for (size_t i = 1; i + 1 < route.size(); ++i) {
|
||||
|
||||
// Try moving point i backward (to positions before i)
|
||||
for (size_t j = 1; j + 2 < i; ++j) {
|
||||
// Current cost: edges around point i and edge j→(j+1)
|
||||
double curLen = dist(pts[route[i - 1]], pts[route[i]])
|
||||
+ dist(pts[route[i]], pts[route[i + 1]])
|
||||
+ dist(pts[route[j]], pts[route[j + 1]]);
|
||||
|
||||
// New cost: bypass i, insert i after j
|
||||
double newLen = dist(pts[route[i - 1]], pts[route[i + 1]])
|
||||
+ dist(pts[route[j]], pts[route[i]])
|
||||
+ dist(pts[route[i]], pts[route[j + 1]]) + 1e-5;
|
||||
|
||||
if (newLen < curLen) {
|
||||
// Move point i to position after j
|
||||
int node = route[i];
|
||||
route.erase(route.begin() + i);
|
||||
route.insert(route.begin() + j + 1, node);
|
||||
relocateFound = true;
|
||||
improvementFound = true;
|
||||
}
|
||||
}
|
||||
|
||||
// Try moving point i forward (to positions after i)
|
||||
for (size_t j = i + 1; j + 1 < route.size(); ++j) {
|
||||
double curLen = dist(pts[route[i - 1]], pts[route[i]])
|
||||
+ dist(pts[route[i]], pts[route[i + 1]])
|
||||
+ dist(pts[route[j]], pts[route[j + 1]]);
|
||||
|
||||
double newLen = dist(pts[route[i - 1]], pts[route[i + 1]])
|
||||
+ dist(pts[route[j]], pts[route[i]])
|
||||
+ dist(pts[route[i]], pts[route[j + 1]]) + 1e-5;
|
||||
|
||||
if (newLen < curLen) {
|
||||
int node = route[i];
|
||||
route.erase(route.begin() + i);
|
||||
route.insert(route.begin() + j, node);
|
||||
relocateFound = true;
|
||||
improvementFound = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// ========================================================================
|
||||
// STEP 4: Remove temporary start/end points
|
||||
// ========================================================================
|
||||
// The temporary markers served their purpose during optimization.
|
||||
// Now remove them so they don't appear in the final result.
|
||||
if (tempEndIdx != -1 && !route.empty() && route.back() == tempEndIdx) {
|
||||
route.pop_back();
|
||||
}
|
||||
if (tempStartIdx != -1 && !route.empty() && route.front() == tempStartIdx) {
|
||||
route.erase(route.begin());
|
||||
}
|
||||
|
||||
// ========================================================================
|
||||
// STEP 5: Map route indices back to original point array
|
||||
// ========================================================================
|
||||
// Since we inserted a temp start point at index 0, all subsequent indices
|
||||
// are offset by 1. Adjust them back to match the original points array.
|
||||
std::vector<int> result;
|
||||
for (int idx : route) {
|
||||
// Adjust for temp start offset
|
||||
if (tempStartIdx != -1) {
|
||||
--idx;
|
||||
}
|
||||
// Only include valid indices from the original points array
|
||||
if (idx >= 0 && idx < static_cast<int>(points.size())) {
|
||||
result.push_back(idx);
|
||||
}
|
||||
}
|
||||
return result;
|
||||
}
|
||||
} // namespace
|
||||
|
||||
@@ -64,194 +301,11 @@ void twoOptSwap(std::vector<int>& path, size_t i, size_t k)
|
||||
* - If both are provided, the path will respect both constraints while optimizing the middle path
|
||||
* - The algorithm ensures all points are visited exactly once
|
||||
*/
|
||||
|
||||
|
||||
std::vector<int> TSPSolver::solve(const std::vector<TSPPoint>& points,
|
||||
const TSPPoint* startPoint,
|
||||
const TSPPoint* endPoint)
|
||||
{
|
||||
size_t n = points.size();
|
||||
if (n == 0) {
|
||||
return {};
|
||||
}
|
||||
|
||||
// Start with a simple nearest neighbor path
|
||||
std::vector<bool> visited(n, false);
|
||||
std::vector<int> path;
|
||||
|
||||
// If startPoint provided, find the closest point to it
|
||||
size_t current = 0;
|
||||
if (startPoint) {
|
||||
double minDist = std::numeric_limits<double>::max();
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
double d = dist(points[i], *startPoint);
|
||||
if (d < minDist) {
|
||||
minDist = d;
|
||||
current = i;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
path.push_back(static_cast<int>(current));
|
||||
visited[current] = true;
|
||||
for (size_t step = 1; step < n; ++step) {
|
||||
double min_dist = std::numeric_limits<double>::max();
|
||||
size_t next = n; // Use n as an invalid index
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
if (!visited[i]) {
|
||||
double d = dist(points[current], points[i]);
|
||||
if (d < min_dist) {
|
||||
min_dist = d;
|
||||
next = i;
|
||||
}
|
||||
}
|
||||
}
|
||||
current = next;
|
||||
path.push_back(static_cast<int>(current));
|
||||
visited[current] = true;
|
||||
}
|
||||
|
||||
// 2-Opt optimization
|
||||
bool improved = true;
|
||||
while (improved) {
|
||||
improved = false;
|
||||
for (size_t i = 1; i < n - 1; ++i) {
|
||||
for (size_t k = i + 1; k < n; ++k) {
|
||||
double delta = dist(points[path[i - 1]], points[path[k]])
|
||||
+ dist(points[path[i]], points[path[(k + 1) % n]])
|
||||
- dist(points[path[i - 1]], points[path[i]])
|
||||
- dist(points[path[k]], points[path[(k + 1) % n]]);
|
||||
if (delta < -Base::Precision::Confusion()) {
|
||||
twoOptSwap(path, i, k);
|
||||
improved = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Handle end point constraint if specified
|
||||
if (endPoint) {
|
||||
// If both start and end points are specified, we need to handle them differently
|
||||
if (startPoint) {
|
||||
// Find the closest points to start and end
|
||||
size_t startIdx = 0;
|
||||
size_t endIdx = 0;
|
||||
double minStartDist = std::numeric_limits<double>::max();
|
||||
double minEndDist = std::numeric_limits<double>::max();
|
||||
|
||||
// Find the indices of the closest points to both start and end points
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
// Find closest to start
|
||||
double dStart = dist(points[i], *startPoint);
|
||||
if (dStart < minStartDist) {
|
||||
minStartDist = dStart;
|
||||
startIdx = i;
|
||||
}
|
||||
|
||||
// Find closest to end
|
||||
double dEnd = dist(points[i], *endPoint);
|
||||
if (dEnd < minEndDist) {
|
||||
minEndDist = dEnd;
|
||||
endIdx = i;
|
||||
}
|
||||
}
|
||||
|
||||
// If start and end are different points, create a new path
|
||||
if (startIdx != endIdx) {
|
||||
// Create a new path starting with the start point and ending with the end point
|
||||
// This ensures both constraints are met
|
||||
std::vector<bool> visited(n, false);
|
||||
std::vector<int> newPath;
|
||||
|
||||
// Add start point
|
||||
newPath.push_back(static_cast<int>(startIdx));
|
||||
visited[startIdx] = true;
|
||||
|
||||
// Add all other points except end point using nearest neighbor algorithm
|
||||
// This builds a path that starts at startIdx and visits all intermediate points
|
||||
size_t current = startIdx;
|
||||
for (size_t step = 1; step < n - 1; ++step) {
|
||||
double minDist = std::numeric_limits<double>::max();
|
||||
size_t next = n; // Invalid index (n is out of bounds)
|
||||
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
if (!visited[i] && i != endIdx) {
|
||||
double d = dist(points[current], points[i]);
|
||||
if (d < minDist) {
|
||||
minDist = d;
|
||||
next = i;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (next == n) {
|
||||
break; // No more points to add
|
||||
}
|
||||
|
||||
current = next;
|
||||
newPath.push_back(static_cast<int>(current));
|
||||
visited[current] = true;
|
||||
}
|
||||
|
||||
// Add end point as the final stop in the path
|
||||
newPath.push_back(static_cast<int>(endIdx));
|
||||
|
||||
// Apply 2-opt optimization while preserving the start and end points
|
||||
// The algorithm only swaps edges between interior points
|
||||
bool improved = true;
|
||||
while (improved) {
|
||||
improved = false;
|
||||
// Start from 1 and end before the last point to preserve start/end constraints
|
||||
for (size_t i = 1; i < newPath.size() - 1; ++i) {
|
||||
for (size_t k = i + 1; k < newPath.size() - 1; ++k) {
|
||||
// Calculate improvement in distance if we swap these edges
|
||||
double delta = dist(points[newPath[i - 1]], points[newPath[k]])
|
||||
+ dist(points[newPath[i]], points[newPath[k + 1]])
|
||||
- dist(points[newPath[i - 1]], points[newPath[i]])
|
||||
- dist(points[newPath[k]], points[newPath[k + 1]]);
|
||||
|
||||
// If the swap reduces the total distance, make the swap
|
||||
if (delta < -Base::Precision::Confusion()) {
|
||||
std::reverse(newPath.begin() + static_cast<long>(i),
|
||||
newPath.begin() + static_cast<long>(k) + 1);
|
||||
improved = true;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
path = newPath;
|
||||
}
|
||||
// If start and end are the same point, keep path as is
|
||||
}
|
||||
else {
|
||||
// Only end point specified (no start point constraint)
|
||||
// Find the point in the current path that's closest to the desired end point
|
||||
double minDist = std::numeric_limits<double>::max();
|
||||
size_t endIdx = 0;
|
||||
for (size_t i = 0; i < n; ++i) {
|
||||
double d = dist(points[path[i]], *endPoint);
|
||||
if (d < minDist) {
|
||||
minDist = d;
|
||||
endIdx = i;
|
||||
}
|
||||
}
|
||||
|
||||
// Rotate the path so that endIdx is at the end
|
||||
// This preserves the relative order of points while ensuring the path ends
|
||||
// at the point closest to the specified end coordinates
|
||||
if (endIdx != n - 1) {
|
||||
std::vector<int> newPath;
|
||||
// Start with points after endIdx
|
||||
for (size_t i = endIdx + 1; i < n; ++i) {
|
||||
newPath.push_back(path[i]);
|
||||
}
|
||||
// Then add points from beginning up to and including endIdx
|
||||
for (size_t i = 0; i <= endIdx; ++i) {
|
||||
newPath.push_back(path[i]);
|
||||
}
|
||||
path = newPath;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return path;
|
||||
return solve_impl(points, startPoint, endPoint);
|
||||
}
|
||||
|
||||
@@ -626,7 +626,8 @@ def sort_locations_tsp(locations, keys, attractors=None, startPoint=None, endPoi
|
||||
- endPoint: Optional ending point [x, y]
|
||||
|
||||
Returns the sorted list of locations in TSP order.
|
||||
If startPoint is None, path starts from the first point in the original order.
|
||||
If startPoint is None, the path is optimized to start near the first point in the original list,
|
||||
but may not start exactly at that point.
|
||||
"""
|
||||
# Extract points from locations
|
||||
points = [(loc[keys[0]], loc[keys[1]]) for loc in locations]
|
||||
|
||||
Reference in New Issue
Block a user