Compact Routes for the Min-Max K Windy Rural Postman Problem by Oliver Lum1, Carmine Cerrone2, Bruce Golden3, Edward Wasil4 1. Department of Applied Mathematics and Scientific Computation, University of Maryland, College Park 2. Department of Computer Science, University of Salerno 3. R.H Smith School of Business, University of Maryland, College Park 4 Kogod School of Business, American University 2 Problem Motivation The MMKWRPP A natural extension of the Windy Rural Postman Problem Route 2 = Required = Included in route = Not traversed Route 1 • • • • • Minimize the max route cost Homogenous fleet of K vehicles Asymmetric traversal costs Required and unrequired edges Generalization of the directed, undirected, and mixed variants • Takes into account route balance and customer satisfaction Depot Route 3 3 Generality One of the most appealing features of the Min-Max K Windy Rural Postman Problem is that it has many fundamental arc routing problems as special cases. MMKWRPP Graph Transformation MMKURPP MMKDRPP MMKMRPP Single-Vehicle URPP DRPP MRPP WRPP Full-Service P CPP DCPP MCPP MCPP 4 Literature Review The MMKWRPP 1 2 3 Introduction Metaheuristic Exact Solver Benavent, Enrique, et al. “Minmax k-vehicles windy rural postman problem.” Networks 54:4 (2009): 216-226. Benavent, Enrique, Angel Corberan, Jose M. Sanchis. “A metaheuristic for the min-max windy rural postman problem with k vehicles.” Computational Management Science 7:3 (2010): 269-287. Benavent, Enrique, et al. “A branch-price-and-cut method for the min-max k-windy rural postman problem.” Networks 63:1 (2014): 34-45. . • ILP Formulation • Polyhedron Characterized • Valid Inequalities (Aggregated, Disaggregated, R-odd cut, Honeycomb, etc.) • Route-First, Cluster-Second Heuristic • Multi-Start, ILS Metaheuristic based on single-vehicle work by same authors • Improves on the 2009 work • Adds pricing scheme • Faster, more scalable method, used to solve larger instances 5 Algorithm of Benavent et al. The MMKWRPP Step 1: WRPP 4 Solve the single-vehicle variant. Step 2: Compact Route Representation 2 5 3 This produces a solution that can be represented as an ordered list of required edges (where any gaps are traversed via shortest paths) 6 1 Step 3: Split Solve for the optimal split of the route into k distinct routes, by finding k-1 points in the route to return to the depot, preserving ordering Depot 8 7 6 Algorithm of Benavent et al. The MMKWRPP • Construct a directed, acyclic graph (DAG) with m+1 vertices, (0,1,…,m) where the cost of the arc (i-1,j) is the cost of the tour starting at the depot, going to the tail of edge i, continuing along the single-vehicle solution through edge j, and then returning to the depot 7 Algorithm of Benavent et al. The MMKWRPP 4 2 5 3 6 1 Depot 8 7 8 Algorithm of Benavent et al. The MMKWRPP • Find a k-edge narrowest path (a path in which the weight of the heaviest edge in the traversal is minimized) from v0 to vm in the DAG, corresponding to a solution • A simple modification to Dijkstra’s single-source shortest path algorithm can produce such a path 9 Algorithm of Benavent et al. The MMKWRPP Step 1: WRPP 4 Solve the single-vehicle variant.. Step 2: Compact Route Representation 2 This produces a solution that can be represented as an ordered list of required edges (where any gaps are traversed via shortest paths). 5 3 x x 6 1 Step 3: Split Solve for the optimal split of the route into k distinct routes, by finding k-1 points in the route to return to the depot, preserving ordering Depot 8 7 10 Algorithm of Benavent et al. The MMKWRPP B C A A={red, yellow} B={black, blue, teal} C={black, yellow, teal} 11 Partitioning Approach The MMKWRPP • Transform the graph into a vertex-weighted graph by constructing its edge dual in the following way: • Create a vertex for each edge in the original graph • Connect two vertices I and j if, in the original graph, edge I and edge j shared an endpoint 4 2 3 5 7 1 6 Depot 12 Partitioning Approach The MMKWRPP • Set the vertex weights to account for known dead-heading and distance to the depot 4 2 5 3 6 7 1 Depot dist (i ) wi c(i ) * (1 d i ) 2 * nearest (i ) |E| |R| ì0 di = í î1 if link i must be deadheaded otherwise 13 Partitioning Approach The MMKWRPP • Partition the transformed graph into k approximately equal parts 4 2 5 3 7 6 1 Depot Green Edge Green Vertex 14 Partitioning Approach The MMKWRPP • For each of the partitions, solve the single-vehicle problem for which only the required edges in the partition are actually required 4 4 4 2 2 2 5 3 1 Depot 5 3 7 7 6 5 3 6 1 Depot 7 6 1 Depot 15 Partitioning Approach The MMKWRPP • Visually appealing • Customers on the same route are close to each other • Other than travel to and from the depot, little overlap • Routes further from depot are smaller • Customers as contiguous as possible 16 Comparing Partitions The MMKWRPP 17 Aesthetic Measures • In practice, routes often exhibit properties like connectedness and compactness • Two metrics (ROI, ATD) proposed in Constantino et al. (European Journal of Operational Research, 2015) are the first to feature interactions between routes • We introduce a third metric, Hull Overlap (HO), that incorporates the intuition behind ROI and ATD Hull Overlap Route Overlap Index NO | N | ROI 2 ( | R | | N | 1) | N | | R| NO n iN r 1 r i Average Traversal Distance | R| ATD Dist r 1 a ,bS r taskpairs ab 18 Route Overlap Index (ROI) Compactness Metrics Attempts to measure the degree to which a set of routes overlaps. It penalizes each ‘required’ node for every route in which it’s visited, and normalizes based on an ‘ideal’, square instance (shown below on the right) Formula Motivation NO | N | ROI 2 ( | R | | N | 1) | N | | R| NO n iN r 1 r i Route Overlap Index Node Overlap Square Instance Square Routes Border Compensation 19 Average Traversal Distance (ATD) Compactness Metrics Attempts to measure the compactness of a set of routes. It penalizes pairwise shortest path distances between links requiring service. Formula Motivation | R| ATD Dist r 1 a ,bS r ab Depot taskpairs taskpairs Non-compact Routes tasks * ( tasks R ) Average Traversal Distance 2* R Pairwise Dist. Depot Compact Routes Task Pairs Non-Comp. Routes Compact Routes 20 Hull Overlap (HO) Compactness Metrics Attempts to measure the degree to which a set of routes overlaps. It calculates the average portion of a route that overlaps with others. Formula Motivation Set of routes in the solution Area of the intersection of arguments Convex hull of the points comprising the argument Depot Area of the argument First Process Second Process Non-compact Routes Third Process Fourt Process Final Process 21 Computational Results The MMKWRPP 10 real street networks taken from cities using the crowd-sourced Open Street Networks database 10 artificial rectangular networks, with random costs between 1 and 10 Experiments run with 3, 5, and 10 vehicles, with 20%, 50%, and 80% of links required Test Specs: • 64-bit PC • Intel i5 4690K 3.5 GHz CPU • 8 GB RAM Metrics: • Distance of longest route • Average Traversal Distance • Route Overlap Index • Hull Overlap 22 Computational Results on Real Street Networks The MMKWRPP • 60 test instances (3 fleet size variations, and 2 depot locations for each of the 10 underlying networks) • |V| ranges from 506 to 2027 • |E| ranges from 586 to 2588 • With respect to max distance, BENAVENT outperforms LUM by 2.36% on average • With respect to ROI, LUM outperforms BENAVENT by 81.7% on average • With respect to ATD, LUM outperforms BENAVENT by 22.9% on average • With respect to HO, LUM outperforms BENAVENT by 26.8% on average • BENAVENT runs into memory constraints on the largest two networks. Results only consider the 48 instances both approaches were able to solve 23 Computational Results on Artificial Networks The MMKWRPP • 60 test instances (3 fleet size variations, and 2 depot locations for each of the 10 underlying networks) • |V| ranges from 225 to 576 • |E| ranges from 420 to 1104 • With respect to max distance, BENAVENT outperforms LUM by 4.38% on average • With respect to ROI, LUM outperforms BENAVENT by 72.7% on average • With respect to ATD, LUM outperforms BENAVENT by 29.6% on average • With respect to HO, LUM outperforms BENAVENT by 38.6% on average 24 Conclusions The MMKWRPP • In practice, many routing problems require visually appealing Route Optimize a solutions Refine the Quality Multi-Objective Partitions • We reviewed previous attempts in the literature to Function quantify what Survey constitutes a ‘visually appealing’ set of routes and proposed our own metric that captures additional intuition • We presented an algorithm to solve a general arc routing variant and compared solutions with the existing state-of-the-art procedure • We showed the tradeoff between performance with respect to the objective function and the aesthetic quality of the routes • Computational results demonstrate consistent relative performance, robust to network layout, fleet size, and depot position 25 Future Work The MMKWRPP Refine the Partitions Improvement procedures and transformations to iteratively alter the partition Route Quality Survey Optimize a Multi-Objective Function Verify and motivate new metric design based on the results of what practitioners actually consider ‘goodlooking’ routes Build the new metrics into the optimization procedures so that it’s possible to tune a solution technique to the relative importance of having aesthetically pleasing routes