Core area territory planning for optimizing driver familiarity and route flexibility

[0001] This application is a divisional application of U.S. patent application Ser. No. 11/962,871 filed Dec. 21, 2007 now U.S. Pat No. 7,660,651, which is a divisional application of U.S. patent application Ser. No. 10/647,062 filed Aug. 22, 2003 now U.S. Pat. No. 7,363,126, which claims the benefit and priority of the U.S. Provisional Patent Application bearing Ser. No. 60/405,138, filed Aug. 22, 2002, all of which are hereby incorporated herein in their entirety by reference.

[0002] The planning of efficient delivery routes presents a variety of technical and logistical challenges. Routes may be fixed or variable. Postal mail delivery is an example of a fixed-route system, in which deliveries are made to the same addresses each day along a fixed route. Variable routes greatly complicate the route planning process. In many modern service businesses, delivery routes change in order to accommodate the changing needs of customers. In a parcel delivery service business, for example, delivery and pickup routes change frequently, as often as every day. Incremental improvements in efficiency within a variable-route system can yield great dividends in performance, profitability, and overall value.

[0003] A region to be served by a business may be divided into a collection of local service territories. Each service territory may include one or more hubs, from which a staff of service providers (drivers) in a fleet of vehicles may be dispatched to serve the territory. For many types of businesses, customer participation and daily demand are generally stochastic (random). A subset of customers with a repeating or daily need may sometimes be identified. Generally, however, the list of participating customer addresses will vary significantly on any given day. The types of services provided along the route may also vary significantly. For example, the service may include pickups as well as deliveries. Additionally, the service types often include specific pickup times or guaranteed delivery times.

[0004] The system constraints on a service business include the number and capacity of the vehicles in its fleet, the number of drivers, and the number of hours in a work day. The geography of the service territory also creates a unique set of constraints and challenges.

[0005] One approach to serving a territory, for example, may include dispatching vehicles from a central hub to a specific outlying area or cluster. The cluster method is not the most efficient approach for several reasons, including the fact that each vehicle is not being fully utilized between the hub and the outlying cluster. The inefficiencies of the cluster method increase as a service territory grows in size.

[0006] Another approach sometimes used for a service territory, for example, is to dispatch vehicles along major roads extending away from the hub, and along the same or different major roads when returning to the hub at the end of the work day. While this practice of dividing the territory into loop routes may meet customer demand, it is typically not the most optimal way to service an area. Like the cluster method, the loop method becomes less and less efficient as the service territory grows in size and the loops become narrower and longer.

[0007] Other approaches developed based on experience or tradition also become inefficient over time and result in significant losses in productivity when a service territory grows in size and complexity.

[0008] A variety of mathematical solutions have been developed in the field which attempt to generate solutions for the so-called Vehicle Routing Problem (VRP). These complex algorithms are generally impractical for daily route planning because they typically require far too much computation time. Add the variable of stochastic customer demand, and the mathematical algorithms grow more complex and require even longer computation times. Thus, there exists a need in the art for a route planning and optimization system capable of generating a solution quickly enough to dispatch a fleet of vehicles on a daily basis.

[0009] Driver familiarity with particular routes and customers is a goal for many service businesses. To date, however, the factor of driver familiarity has not been considered when planning daily routes that are optimized on a daily basis to meet a variable demand. Thus, there exists a need in the art for a route planning and optimization system that promotes driver familiarity while meeting the efficiency requirements of a stochastic demand.

[0010] The following summary is not an extensive overview and is not intended to identify key or critical elements of the matter disclosed, nor is it intended to delineate the scope of such elements. This Summary provides a conceptual introduction in a simplified form as a prelude to the more-detailed description that follows.

[0011] In one aspect, a method is provided. In one embodiment, the method includes (a) selecting a new driver from a pool of unassigned drivers; (b) identifying a plurality of cells, wherein each of the plurality of cells comprises (i) a portion of a service territory and (ii) a group of one or more stops; (c) classifying one or more of the plurality of cells as core cells in response to a known service volume for each of the core cells exceeding a minimum service volume; and (d) establishing one or more core areas, wherein each of the one or more core areas comprises a localized cluster of the one or more core cells. The method also includes (e) selecting a new core area from the one or more core areas based upon a known driver visit frequency by the new driver to each of the one or more core areas; (f) assigning the new driver to service the new core area after selecting the new core area; and (g) classifying the new driver as an assigned driver after assigning the new driver to service the new core area.

[0012] In accordance with another aspect, a computer program product is provided. The computer program product may comprise at least one computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising executable portions configured to (a) select a new driver from a pool of unassigned drivers; (b) identify a plurality of cells, wherein each of the plurality of cells comprises (i) a portion of a service territory and (ii) a group of one or more stops; (c) classify one or more of the plurality of cells as core cells in response to a known service volume for each of the core cells exceeding a minimum service volume; and (d) establish one or more core areas, wherein each of the one or more core areas comprises a localized cluster of the one or more core cells. The computer-readable program code portions may also comprise executable portions configured to (e) select a new core area from the one or more core areas based upon a known driver visit frequency by the new driver to each of the one or more core areas; (f) assign the new driver to service the new core area after selecting the new core area; and (g) classify the new driver as an assigned driver after assigning the new driver to service the new core area.

1. Introduction

[0038] Exemplary systems, methods, and apparatuses are now described with reference to the drawing figures, where like reference numerals are used to refer to like elements throughout the several views. In the following description, for purposes of explanation, numerous specific details are set forth in order to facilitate a thorough understanding of the systems, methods, apparatuses, and the like. It may be evident, however, that the exemplars described may be practiced without these specific details. In other instances, common structures and devices are shown in block diagram form in order to simplify the description.

[0039] As used in this application, the term “computer component” refers to a computer-related entity, either hardware, firmware, software, a combination thereof, or to software in execution. For example, a computer component can be, but is not limited to being, a process running on a processor, a processor itself, an object, an executable, a thread of execution, a program, and a computer. By way of illustration, both an application running on a server and the server itself can be a computer component. One or more computer components cans reside within a process and/or thread of execution and a computer component can be localized on a single computer and/or distributed between and among two or more computers.

[0040] “Software,” as used herein, includes but is not limited to, one or more computer readable and/or executable instructions that cause a computer, computer component and/or other electronic device to perform functions, actions and/or behave in a desired manner. The instructions may be embodied in various forms like routines, algorithms, modules, methods, threads, and/or programs. Software may also be implemented in a variety of executable and/or loadable forms including, but not limited to, a stand-alone program, a function call (local and/or remote), a servelet, an applet, instructions stored in a memory, part of an operating system or browser, and the like. It is to be appreciated that the computer readable and/or executable instructions can be located in one computer component and/or distributed between two or more communicating, co-operating, and/or parallel-processing computer components and thus can be loaded and/or executed in serial, parallel, massively parallel and other manners. It will be appreciated by one of ordinary skill in the art that the form of software may be dependent on, for example, requirements of a desired application, the environment in which it runs, and/or the desires of a designer or programmer or the like.

[0041] “Data store,” as used herein, refers to a physical and/or logical entity that can store data. A data store may be, for example, a database, a table, a file, a list, a queue, a heap, and so on. A data store may reside in one logical and/or physical entity and/or may be distributed between two or more logical and/or physical entities.

[0042] The systems, methods, apparatuses, and objects described herein may be stored, for example, on a computer readable media. Media may include, but are not limited to, an ASIC, a CD, a DVD, a RAM, a ROM, a PROM, a disk, a carrier wave, a memory stick, and the like. Thus, an example computer readable medium can store computer executable instructions for a method for managing transportation assets. The method includes planning a route for a transportation asset based on the analysis of data retrieved from an experience-based route database.

[0043] To the extent that the term “includes” is employed in the detailed description or the claims, it is intended to be inclusive in a manner similar to the term “comprising” as that term is interpreted when employed as a transitional word in a claim. Further still, to the extent that the term “or” is employed in the claims (for example, A or B) it is intended to mean “A or B or both.” When the author intends to indicate “only A or B but not both,” the author will employ the phrase “A or B but not both.” Thus, use of the term “or” herein is the inclusive use, not the exclusive use. See Bryan A. Garner, A Dictionary Of Modern Legal Usage 624 (2d ed. 1995).

[0044] It will be appreciated that some or all of the processes and methods of the system involve electronic and/or software applications that may be dynamic and flexible processes so that they may be performed in other sequences different than those described herein. It will also be appreciated by one of ordinary skill in the art that elements embodied as software may be implemented using various programming approaches such as machine language, procedural, object oriented, and/or artificial intelligence techniques.

[0045] The processing, analyses, and/or other functions described herein may also be implemented by functionally equivalent circuits like a digital signal processor circuit, a software controlled microprocessor, or an application specific integrated circuit. Components implemented as software are not limited to any particular programming language. Rather, the description herein provides the information one skilled in the art may use to fabricate circuits or to generate computer software to perform the processing of the system. It will be appreciated that some or all of the functions and/or behaviors of the present system and method may be implemented as logic as defined above.

[0046] Many modifications and other embodiments may come to mind to one skilled in the art who has the benefit of the teachings presented in the description and drawings. It should be understood, therefore, that the invention is not be limited to the specific embodiments disclosed and that modifications and alternative embodiments are intended to be included within the scope of the disclosure and the exemplary inventive concepts. Although specific terms may be used herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

2. A Service Territory

[0047] FIG. 1 illustrates a service territory 20 having a central hub 30 and a plurality of discrete delivery areas referred to herein as cells 40. In this context, a cell 40 can be thought of as the minimum unit of a service territory 20 within which all the customers or stops 42 may be serviced by a single driver. Within a cell 40, as depicted in FIG. 2, a travel route may include a sub-route 45 between and among one or more stops 42 within the cell 40. Each stop 42 may include one or more services, such as a parcel delivery or pickup.

[0048] In one embodiment, the method of the present invention involves grouping nearby stops 42 into cells 40. Stops 42 referred to as being localized are sufficiently close together to form a cell 40. As a general rule, stops 42 are located closely enough together to form a cell 40 if the cell can be serviced completely by a single driver during a workday. In a practical application, a service territory 20 may be partitioned into cells 40 by neighborhood, office building, industrial park, by ZIP+2 or ZIP+4 codes, for example, or by any other manageable unit.

[0049] Gathering customers into cells 40 both simplifies the route planning process and more accurately represents the actual driver experience of learning about particular cells 40 within a service territory 20. In the context of a parcel delivery service, a service territory 20 may be served using a route plan that includes one or more travel routes.

[0050] Planning daily routes to a changing set of participating customers in a service territory 20 is a task sometimes referred to in the field as a random or stochastic Vehicle Routing Problem (VRP). The methods for planning daily routes disclosed herein include modifications and variations of the stochastic VRP to produce timely and useful solutions.

3. Workload within a Cell

[0051] Because a cell 40 is the minimum unit of work assigned to each driver, the approximation of the workload within a cell 40 is needed. Customer demand can be described in terms of a known customer set where: (1) each customer has a probability of participating on a particular day; and (2) the volume of the demand varies each day; for example, in terms of the number of parcel. In the context of a parcel delivery service, the identity of each participating customer and the demand volume for the day is generally learned before the route plans are generated each day.

[0052] In one aspect of the invention, a method for approximating the workload within a cell 40 is accomplished, generally, by solving the traditional Traveling Salesman Problem (TSP) within each cell 40 when the customer location and demand becomes known. The workload approximation model may be expressed as:
W_i=n_is_i+(TSP_i−L_max)/ν_i
where: W_i is the workload within a cell 40;

[0053] n_i is the number of stops 42 in a cell 40;

[0054] s_i is the mean time duration for each stop 42;

[0055] TSP_i is the total distance traveled within a cell 40;

[0056] L_max is the distance of the longest arc on the optimal TSP route; and

[0057] v_i is the average vehicle velocity along the route.

The element n_is_i in the workload model may be referred to as a stop time factor. The element (TSP_i−L_max)/ν_i may be referred to as a travel time factor.

[0058] In one embodiment, the method of the present invention places a higher priority on computing time than on solution quality when solving the Traveling Salesman Problem (TSP) in this context. Quicker results may greatly improve computation efficiency, whereas the delay of obtaining highly-accurate results may harm efficiency. Because quick computing times are desired, the method incorporates a procedure called a Convex Hull heuristic algorithm

[0059] To apply the Convex Hull heuristic algorithm, the following steps may be followed:

• • [0060] (1) Create the convex hull of all the stops 42 within a cell 40. The convex hull of a set of points is defined as the smallest convex polygon that encloses all the points. In this procedure, the convex hull may form an outer loop of hull stops k which, when connected, produce a sub-route 45 within the cell 40. The sub-route 45 of hull stops k serves as a starting point for the computation.
  • [0061] (2) For each stop 42 that is not a hull stop k, find the best insertion position (between stop i and stop j), such that (d_ik+d_kj−d_ij) is minimal, where d_ik is the distance between stop i and k, etc.
  • [0062] (3) For each solution (a sequence of stops: i, k, j) found in step 2, select the best insertion stop k* and its best insertion position (between stop i* and stop j*), such that (C_i*k*+C_k*j*)/C_i*j* is minimal.
  • [0063] (4) Insert stop k* into the sub-route 45 between stop i* and stop j*.
  • [0064] (5) Repeat steps 2, 3, and 4, until all the stops 42 in the sub-route 45 are ordered such that each stop 42 is visited exactly once.

[0065] The Convex Hull heuristic algorithm produces an optimal TSP sub-route within a cell 40. The distances between each stop dik may be summed to find the total distance TSP_i traveled within the cell 40. Furthermore, because each segment between stops is known, the algorithm may also be used to determine the distance L_max of the longest arc on the optimal TSP sub-route. These values, TSP_i and L_max may be used to solve the workload approximation model and approximate the workload W_i within a cell 40.

[0066] The Convex Hull heuristic algorithm as well as the workload approximation equation may be solved by a computer system, programmed to execute the method.

4. Cell-to-Cell Travel Distance

[0067] The cell-to-cell travel distance plays a role in daily route planning because it affects the cost of inserting a stop along a route. The first stop in a cell may be called the entry stop; the last stop may be called the exit stop. The distance from cell i to cell j is expressed as d_ij and it represents the distance from the exit stop in cell i to the entry stop in cell j.

[0068] Instead of using the centroid-to-centroid distance between cells to estimate the cell-to-cell distance, one embodiment of the present invention includes the following distance model:

[0069] $d_{\mathrm{ij}}=b_{\mathrm{ij}}+a_1\frac{\sqrt{a_i}}{\sqrt{n_i}}+a_2\frac{\sqrt{a_j}}{\sqrt{n_j}}$
where: d_ij is the estimated cell-to-cell distance;

[0070] b_ij is the closest boundary distance between cell i and cell j;

[0071] n_i is the number of stops in cell i; and

[0072] a_i is the geographic area of cell i.

[0073] As the number of stops n_i in each cell increases, the distance between adjacent stops decreases; and, thus, the two additive elements in the equation will decrease and approach zero. Also, as the number of stops n_i increases and fills the cell, the stops become located nearer to the cell boundary; and, thus, the cell-to-cell distance d_ij will decrease and approach the boundary distance b_ij.

[0074] The coefficients α₁, α₂ in the second and third additive elements of the equation represent the average travel distance between two stops in each cell, based upon a continuous Traveling Salesman Problem estimation model. Values for the coefficients α₁, α₂ may be estimated from actual route data, using an empirical simulation method. The values for the coefficients α₁, α₂ may then be used to solve the cell-to-cell distance equation.

5. The Value of Driver Familiarity

[0075] Driver familiarity is optimal when a driver travels the same route every day. But in stochastic systems were customers and demand change daily, the routes also change daily.

[0076] In one aspect, the route planning systems disclosed herein provide an estimate of the value of driver familiarity. Although the general importance of driver familiarity may have been recognized qualitatively, the present invention provides a quantitative analysis of driver familiarity in the context of route planning. In the planning systems disclosed, the concepts of the learning curve, the forgetting curve, and a dynamic learning function are applied to the phenomenon of driver familiarity. The route planning systems disclosed also include and incorporate an expected driver performance level for a particular visit frequency to a cell.

[0077] 5.1. The Learning Curve

[0078] A learning curve describes the improvement by a person who performs a task repetitively. As the number of attempts or cycles increases, the overall cost per cycle decreases. In the context of route planning, a driver's performance may be approximated in terms of the time required to finish the workload within a given cell. In one aspect of the present invention, driver performance may be directly related to the frequency of visiting the same cell. A high visit frequency to a cell promotes driver familiarity within the cell and, in turn, improves driver performance overall.

[0079] A learning curve model as shown in FIG. 11 may be expressed as:
T_j=T₁j⁻¹
where: T_j is the time to complete the jth attempt or cycle;

[0080] T₁ is the time to complete the first cycle;

[0081] j is the number of cumulative cycles; and

[0082] l is the learning slope, which may be a constant for any particular situation.

[0083] When the number of cumulative cycles j is doubled, the time required will decreased at a constant rate known as a learning rate ρ). In mathematical terms:

[0084] $\mathrm{because}{T}_n=T_1{n}^{-l}$ $\mathrm{and}{T}_{2n}={T_1\left(2n\right)}^{-l},\frac{T_{2n}}{T_n}=\frac{{\left(2n\right)}^{-l}}{n^{-l}}=2^{-l}=\rho$ $\mathrm{and}l=-{\log }_2^{\rho }.$

[0085] Referring again to the learning curve model (T_j=T₁j⁻¹): as the number of cycles of visits j increase, the time to complete each visit T_j decreases and tends to approach a minimum lower limit called T_f. Applying the concept of learning rate ρ and the lower limit T_f, a driver's learning curve may be expressed as:
T_j=max{T₁j⁻¹,T_f}
where: T_j is the time needed per stop to complete the jth visit to the cell;

[0086] T₁ is the standard time needed to service a stop without learning (i.e., the time needed to complete the first visit to a new cell);

[0087] j is the number of cumulative visits; and

[0088] l is the learning slope, which may be a constant for a particular situation. The learning slope l may be determined by the learning rate ρ.

[0089] To determine the time needed per stop T_j in relation to the standard time T₁, divide both sides of the learning curve equation by T₁:

[0090] $\frac{T_j}{T_1}=\max \left\{1·j^{-l},\frac{T_f}{T_1}\right\}$ $\mathrm{If}{T}_j^{\prime }=\frac{T_j}{T_1},T_0=1$ $\mathrm{and}{T}_{\infty }=\frac{T_f}{T_1},\mathrm{then}{T}_j^{\prime }=\max \left\{T_0·j^{-l},T_{\infty }\right\},$
where T_j′ represents the percentage of the standard time T₁ a driver needs to service each stop on the jth visit to the cell. This equation also describes a learning curve, with the same learning slope, but with a different starting level and a learning limit T_∞. The learning limit T_∞ as shown in FIG. 11 reflects the efficiency gained in stop-to-stop traveling time when a driver has visited the cell enough times to develop maximum familiarity.

[0091] In addition to applying the driver learning model to the stops made within a cell, the model may also be used to analyze the travel time between stops, parcel handling at a single stop, and any other discrete task that improves as familiarity increases. For example, if a cell i is assigned to a core area, the average time that a driver spends on each stop within this cell {tilde over (T)}_i should be:
{tilde over (T)}_i=T_∞·T₀+T_∞·S₀
where T₀ is the beginning or standard traveling time per stop, and S₀ is the standard package handling time per stop. If T_i=T₀+S₀, then {tilde over (T)}_i=T_∞· T_i.

[0092] 5.2. Learning with Interruptions and Forgetting

[0093] At the operational level where routes will vary over time, the driver learning process may be interrupted when the driver does not visit a particular cell on a daily route. An interruption may cause a degree of forgetting about the cell that can be approximated. Like the learning curve phenomenon, a forgetting curve may describe the change in performance that occurs when a person ceases performing a task repetitively. The forgetting may be described as a variable regression of what was learned, toward an invariant level or limit of forgetting.

[0094] FIG. 12 illustrates both a learning curve and a forgetting curve, which may take similar forms. The forgetting curve may be expressed as:
{circumflex over (T)}_x={circumflex over (T)}₁x^f
where: {circumflex over (T)}_X is the time for the xth visit of lost experience;

[0095] {circumflex over (T)}₁ is the equivalent time per stop for the first lost experience;

[0096] x is the number of visits to the cell that would accumulate if interruption never occurred; and

[0097] f is the forgetting slope, which may be a constant for a particular situation.

The value of the forgetting slope f may depend on the forgetting rate.

[0098] By equating {circumflex over (T)}₁ (the time per stop required on the forgetting curve) to the average time per stop required to complete a service in the cell for the q_ith visit on the learning curve, the equation becomes:
{circumflex over (T)}₁=T₁q_i^−(i+f)
where i=1 (for the first missed visit).

[0099] To analyze a finite gap or break period from visiting a particular cell, assume that after q_i continuous visits, the learning process is interrupted for a period of time, during which, if there had been no interruption, an additional s_i visits would have been made. The average time required to service a stop in the cell by the end of the break period on the forgetting curve would be:
{circumflex over (T)}_qi+si={circumflex over (T)}₁(q_i+s_i)^f
As shown in FIG. 12, u_i+1 is the amount of equivalent visits remembered at the beginning of visit cycle (i+1), after the interruption in cycle i. Accordingly:

[0100] $u_{i+1}={\left(\frac{{\hat{T}}_1}{T_1}{\left(q_i+s_i\right)}^f\right)}^{-1/l}$

[0101] The average time required to service a stop in the cell for the first visit after the break period in cycle i may be expressed as:
{circumflex over (T)}_1,i+1={circumflex over (T)}₁(u₁₊₁+1)⁻¹.

[0102] 5.3. The Driver Learning Function

[0103] In one embodiment of the present invention, a system may be used to store data describing the number of times a driver visits a particular cell. Using the equations discussed above, along with the stored visit data for each driver and each cell, a dynamic learning function may be constructed, as shown in FIG. 13. The function curve includes interruptions and forgetting g(n) as shown.

[0104] When the number of customers in cell (i, Di) is known, the total workload in terms of travel time and service time for a driver j can be expressed as:
f(n_i,D_i)=D_ig(n_i),
where f (n_i, D_i) may be referred to as the Driver Learning Function.

[0105] The Driver Learning Function provides useful information at the operational level, where core cells and core areas are assigned to particular drivers. Using historical cell visit data and applying the Driver Learning Function, each driver's dynamic performance level relative to each cell may be calculated. The data may reveal, for example, that Driver Smith shows better performance than Driver Jones within cell i. In other words, the data may show that Driver Smith is more familiar with cell i than Driver Jones. Also, applying the workload calculations, Driver Smith would have a lower (and thus, more efficient) total expected workload when servicing cell i than Driver Jones. Cell i, therefore, would likely be assigned to Driver Smith, who is most familiar with cell i and can accomplish the service with less work.

[0106] 5.4. Driver Performance and Visit Frequency

[0107] Planning routes involves predicting driver performance. Optimizing routes involves evaluating actual driver performance in comparison to the predicted performance.

[0108] Assume that a driver visits a cell according to a probability p. After a cell visit, driver performance will increase along the learning curve. When there is no cell visit, driver performance will decrease along the forgetting curve. Assume the learning slope l and the forgetting slope f are equivalent, such that 0<l=f<1. Assuming a current driver performance level of x, the expected subsequent driver performance level may be predicted to be:

[0109] $p·\max \left\{T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f},T_{\infty }\right\}+\left(1-p\right)·\min \left\{T_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f,T_0\right\}$
Assuming an overall expected driver performance level of {umlaut over (x)} produces the following expression, referred to herein as Equation One:

[0110] $p·\max \left\{T_0·{\left({\left(\frac{\stackrel{⋒}{x}}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f},T_{\infty }\right\}+\left(1-p\right)·\min \left\{T_{\infty }·{\left({\left(\frac{\stackrel{⋒}{x}}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f,T_0\right\}=\stackrel{⋒}{x}$
Defining the function, ƒ₀(x):

[0111] $f_0\left(x\right)=p·\max \left\{T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f},T_{\infty }\right\}+\left(1-p\right)·\min \left\{T_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f,T_0\right\}-x$
If there exists a single solution when ƒ₀(x)=0, then the solution represents the expected driver performance level given a visiting frequency of p.

[0112] To determine if the function value is decreasing while driver performance x is increasing, there are four scenarios to consider:

First Scenario:

[0113] $T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}\le T_{\infty }\mathrm{and}{T}_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f\ge T_0:$
Now function ƒ₀(x) becomes:
f₁(x)=p·T_∞+(1−p)·T₀−x
and f₁(x)<0.
Second Scenario:

[0114] $T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}\le T_{\infty }\mathrm{and}{T}_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f<T_0$
Now function ƒ₀(x) becomes:

[0115] $\begin{array}{c}{f}_2\left(x\right)=p·T_{\infty }+\left(1-p\right)·T_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f-x\\ f_2^{\prime }\left(x\right)=\left(1-p\right)·T_{\infty }·f·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^{f-1}·\left(\frac{1}{f}\right)·{\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}-1}·\left(\frac{1}{T_{\infty }}\right)-1\\ \left(1-p\right)·\frac{{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^{f-1}}{{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}\right)}^{f-1}}-1\\ =\left(1-p\right)·\frac{1}{{\left(1+{\left(\frac{T_{\infty }}{x}\right)}^{\frac{1}{f}}\right)}^{1-f}}-1<\left(1-p\right)-1=-p<0\end{array}$
Third Scenario:

[0116] $T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}>T_{\infty }\mathrm{and}{T}_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f<T_0$
Now function ƒ₀(x) becomes:

[0117] $f_3\left(x\right)=p·T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}+\left(1-p\right)·T_0-x$ $\mathrm{and}\text{:}$ $\begin{array}{c}{f}_3^{\prime }\left(x\right)=p·T_0·\left(-f\right)·\frac{\left(-\frac{1}{f}\right)·{\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}-1}·\left(\frac{1}{T_0}\right)}{{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}-1\\ =p·\frac{{\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}-1}}{{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}-1\\ =p·\frac{1}{{\left(\frac{x}{T_0}\right)}^{\frac{1}{f}+1}·{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}-1\\ =p·\frac{1}{{\left({\left(\frac{x}{T_0}\right)}^{\frac{1}{f}}\right)}^{f+1}·{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}-1\\ =p·\frac{1}{{\left(1+{\left(\frac{x}{T_0}\right)}^{\frac{1}{f}}\right)}^{f+1}}-1<p-1<0\end{array}$
Fourth Scenario:

[0118] $T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}>T_{\infty }\mathrm{and}{T}_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f<T_0$
Now function ƒ₀(x) becomes:

[0119] $f_4\left(x\right)=p·T_0·{\left({\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f}+\left(1-p\right)·T_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f-x$ $\begin{array}{c}{f}^{\prime }\left(x\right)=p·T_0·\left(-f\right)·\frac{\left(-\frac{1}{f}\right)·{\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}-1}·\left(\frac{1}{T_0}\right)}{{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}+\\ \left(1-p\right)·T_{\infty }·f·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^{f-1}·\left(\frac{1}{f}\right)·{\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}-1}·\left(\frac{1}{T_{\infty }}\right)-1\\ =p·\frac{{\left(\frac{x}{T_0}\right)}^{-\frac{1}{f}-1}}{{\left({\left(\frac{T_0}{x}\right)}^{\frac{1}{f}}+1\right)}^{f+1}}+\left(1-p\right)·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^{f-1}·{\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}-1}-1\\ =p·\frac{1}{{\left(1+{\left(\frac{x}{T_0}\right)}^{\frac{1}{f}}\right)}^{f+1}}+\left(1-p\right)·\frac{1}{{\left(1+{\left(\frac{x}{T_0}\right)}^{\frac{1}{f}}\right)}^{1-f}}-1<\\ p+\left(1-p\right)-1=0\end{array}$
Thus, the value of function ƒ₀(x) is decreasing as driver performance x is increasing.
Also,
f(T₀)=p·T₀·2^−f+(1−p)·T₀−T₀=p·T₀·(2^−f−1)<0
and
f(T_∞)=p·T_∞+(1−p)·T_∞·2^f−T_∞>p·T_∞(1−p)·T_∞−T_∞=0
which proves there exists a single x value that satisfies Equation One, where x is the expected driver performance level given a visit frequency of p.

[0120] In one embodiment, the present invention includes an efficient algorithm for solving Equation One. Understanding that function value ƒ₀(x) is decreasing, a solution algorithm in one embodiment includes a binary search method to look for the value of x that allows the function value ƒ₀(x) to decrease and approach zero.

[0121] Applying this method of the present invention, an expected performance level w(p) was calculated based upon a visit frequency p of 50% (or 0.5), a first-visit learning level T₀=1 on day one, and a final learning level of 70% (or 0.7). The resulting learning curve is shown in FIG. 14. The corresponding shape of function ƒ₀(x) is shown in FIG. 15.

[0122] FIG. 14 also includes additional learning curves for visit frequencies p of 90% and 10%, respectively. The marked variation in learning level when the visit frequency p is lower confirms the qualitative notion that driver performance declines when driver learning is interrupted.

6. Route Consistency

[0123] The study and quantification of the value of driver familiarity confirms the intrinsic value of consistent delivery routes. A consistent route means that roughly the same geographical area is served by the same driver each day.

[0124] Consistent routes have both tangible and intangible benefits. A driver familiar with a route spends less time finding addresses and offices, and less time researching maps and directions. Better planning of stops along a consistent route will increase productivity; i.e., increase the number of stops and deliveries completed in a workday. Consistent routes also increase loading efficiency because the personnel who load the packages into the vehicles also learn the characteristics of a particular route. Driver familiarity along a consistent route also provides more personalized service to customers along the route.

[0125] 6.1. Empirical Data and the Grid Method

[0126] Existing algorithms for solving a Vehicle Routing Problem (VRP) do not measure whether or how often a particular customer is visited by the same driver. In one embodiment of the present invention, route planning systems are disclosed that include driver consistency in the context of solving a particular type of VRP.

[0127] FIG. 16 is an illustration of part of a service territory 20 in plan view, showing three routes 1, 2, and 3, assigned to three specific drivers A, B, and C, respectively, on a first day. Instead of showing each and every stop along each route in FIG. 16, each route is represented by its convex hull. The convex hull of a set of points is defined as the smallest convex polygon that encloses all points in the set. The convex hulls for routes 1, 2, 3 are shown using solid lines.

[0128] The dotted lines represent a set of new routes 1′, 2′, 3′, 4′ on a second or subsequent day. Each route 1′, 2′, 3′, 4′ must be assigned to one of the drivers A, B, C. The problem solved by one embodiment of the present invention is to assign the routes while maintaining driver consistency as much as possible, and while considering the known constraints of the number and capacity of the vehicles, the number of drivers, and the hours available in a work day.

[0129] In one embodiment, the route planning system 10 of the present invention solves this problem by selecting a reference day on which the minimum number of drivers is available. Each driver is assigned a route on the reference day. For subsequent days, an attempt is first made to assign the same driver to the new route that has the most amount of territory overlap with the assigned route. The following steps illustrate an embodiment of a driver route assignment method:

• • [0130] (1) Execute a route planning algorithm over the service territory 20 for each day during a reference period.
    • [0131] (a) Determine the number of assigned routes needed for each day during the reference period. The number of assigned routes equals the number of drivers needed.
    • [0132] (b) Find the minimum number of drivers (i.e., the base staff) needed during any single day (or days) during the reference period.
  • [0133] (2) Select as a reference day any single day during the reference period on which only the base staff was needed. The solution of the route planning algorithm for the reference day produces an assigned route plan comprising a number of assigned routes equal to the number of base staff. Each assigned route has an assigned driver.
  • [0134] (3) Compare a current day to the reference day. The route planning algorithm run for the current day produces a plurality of unassigned routes.
    • [0135] (a) For each of the assigned routes on the reference day:
      • [0136] (i) find the one of the plurality of unassigned routes for the current day which has the maximum geographic overlap with the assigned route, wherein the geographic overlap is determined by calculating the area of the intersection between the convex hull of the assigned route and the convex hull of the unassigned route; and
      • [0137] (ii) assign the unassigned route to the assigned driver.
    • [0138] (b) Assign an additional driver to any unassigned routes remaining after each of the number of assigned routes has an assigned driver. In other words, additional drivers available on the current day will serve the remaining unassigned routes.

[0139] Other methods for assigning routes and drivers are disclosed herein, including the following grid method.

[0140] 6.1.1. The Grid Method for Measuring Route Consistency

[0141] In one embodiment, the present invention provides a method for measuring route consistency within a service territory. As shown in FIGS. 3A and 4A, the service territory 20 may be been divided into segments using a grid 150.

[0142] The grid method is illustrated beginning with FIG. 3, which depicts an exemplary first daily route 121 assigned to a first driver 23 (Driver Smith) on a day (Monday) during a reference period. The route 121 may represent one of the routes within the convex hull B illustrated in FIG. 16. In this context, the following steps illustrate an embodiment of a grid method.

[0143] (1) A first step may include dividing the service territory 20 or a portion thereof into a plurality of grid segments 152 using a grid 150. The grid 150 may be take any shape and may include grid segments 152 of any size or shape, including random shapes and including various sizes and shapes within the same grid 150. Empirical data, experience, and historical route data may assist a route planner when imposing a grid 150 upon a particular area. A symmetrical and rectilinear grid 150 is shown in FIG. 3A by way of example and not limitation. Note the grid 150 in FIG. 3A has been superimposed upon the first daily route 121 shown in FIG. 3, such that each cell 40 is represented by a corresponding grid segment 152. In a real operational setting, the grid segments 152 need not coincide with the cells 40. In fact, a grid segment 152 may encompass several cells 40, only a portion of a cell or cells, or no active cells at all. The grid shape depends on the particular characteristics of a service territory 20 or portion thereof and the peculiarities of the routes.

[0144] (2) The following steps may be executed for each grid segment 152, for each day during a reference period, and for each assigned driver. It may be appreciated by one skilled in the art that storing the following data in a database or other type of data store on a computer may facilitate the comparisons, calculations, and other uses described for the data. In one embodiment of the present invention, the methods and systems disclosed impliedly include a method of gathering, tracking, storing, and retrieving empirical data from each element in a route planning system, including but not limited to the hubs, route plans, discrete routes, route stops, cells, sub-routes within cells, sub-route stops, parcels, as well as the drivers, vehicles, and related systems involved in accomplishing deliveries. The grid method may be formulated for execution by a computer system. In one embodiment, an array or matrix may be used to categorize and easily reference the data regarding each grid segment 152, for each day during a reference period, and for each assigned driver, in the following steps:

• • [0145] (a) Compare the number of stops within each grid segment 152 (referred to as grid stops) to the number of total stops along the assigned route for each driver. In FIG. 3A, the number of grid stops is shown as a numeral within each grid segment 152 and corresponds to the cells 40 shown in FIG. 3. The comparison of grid stops to total stops may be expressed as a percentage. For mathematical simplicity, the number of total stops in FIG. 3A equals one hundred.
  • [0146] (b) Repeat the comparison described in step (a) for each day during a reference period. A second daily route 122 with different cell locations is shown in FIG. 4. The second daily route 122 is assigned to the first driver 23 (Driver Smith) on a subsequent day (Tuesday) during a reference period. In FIG. 4A, the number of grid stops is shown as a numeral within each grid segment 152 and corresponds to the cells 40 shown in FIG. 4. Again, for mathematical simplicity, the number of total stops in FIG. 4A equals one hundred.

[0147] (3) An exemplary compilation of the plurality of daily routes 118 actually driven by the first driver 23 during an entire reference period is shown in FIG. 5. In this step, for each driver, over all the days in the reference period, compare the average number of grid stops to the average total number of stops. This comparison is referred to as the grid segment visiting frequency 154 and it may be expressed as a percentage. The grid segment visiting frequency 154 for the first driver 23 only is shown in FIG. 5A. Similar data may be calculated for each driver who serviced a route during the reference period.

[0148] (4) Compare the grid segment visiting frequency 154 across the service territory 20 for each driver. In real operating conditions, a grid 150 similar to that shown in FIG. 5A may be produced for each driver on a staff. Then, for each grid segment 152, identify the driver who has the highest grid segment visiting frequency 154 and assign this most frequent driver to service that grid segment 152 on a continuing basis. In other words, assign the most frequent driver to each grid segment 152.

[0149] (5) Finally, to understand the pattern of visits across the service territory 20, a grid consistency index for each grid segment 152 may be calculated by averaging all the grid segment visiting frequencies 154 of all the drivers. A map of the service territory 20 showing the grid consistency index for each grid segment 152 may facilitate the identification of those geographical areas that are most often serviced by a consistent driver. Thus, the grid consistency index provides a measure of route consistency.

[0150] In another aspect of the grid method, the grid consistency index may form the basis of a Core Area Route Design model, as described below.

[0151] 6.2. A Mathematical Model for Route Consistency

[0152] In one embodiment, the present invention provides a method for measuring route consistency within a service territory using a mathematical model. There exists an opposing tension between route consistency and optimal route planning because route optimization requires flexibility (as opposed to consistency). The value of route flexibility at the individual cell level is derived from the ability of the inventive system to assign the cell to any one driver at a nearly equal cost. Until the development of the methods and systems disclosed herein, the cost of assigning a driver has not included the value of driver familiarity.

[0153] In one embodiment of the system, each cell 40 in a service territory 20 may be classified as a core cell 60, a flex zone cell 50 (discussed below), or a daily cell 70 (i.e., unclassified or, in other words, assigned on a daily basis only). Core cells 60 may be grouped into a core area 100. An example of such a cell classification is depicted in FIGS. 7 through 10.

[0154] In a mathematical model of the present invention, all unassigned cells are assigned to a dummy core area zero. Let I_ik represent the assignment that assigns cell i to core area k. Then the total time associated with assignment I_Ik is:
f_ik=g_ik+h_ik
where: g_ik=the total travel time it takes the driver who is serving this core cell to complete the workload within the cell; and

[0155] h_ik=the total cell-to-cell travel time contribution of assignment I_ik in the optimal cell tour that includes core area k.

[0156] An objective is to minimize the expected total durations associated with the assignments of cells to core areas 100, under the probabilistic constraints that limit the probability that the sum of a workload within each core area 100 will exceed the maximum workday duration of drivers. Accordingly, the following mathematical model was developed:

[0157] $\begin{array}{cc}\mathrm{Min}\mathrm{ }E\left[\sum _{i,k}{f}_{\mathrm{ik}}{I}_{\mathrm{ik}}\right]& \left(1^{\prime }\right)\\ S.t.\sum _k{I}_{\mathrm{ik}}=1,i=1,2,...,nk=0,1,...m& \left(2^{\prime }\right)\\ P\left(\sum _{i,k}{f}_{\mathrm{ik}}{I}_{\mathrm{ik}}\le Q_k\right)\ge 1-\alpha ,\left(k=1,...,m\right)& \left(3^{\prime }\right)\\ I_{\mathrm{ik}}=0\mathrm{or}1,i=1,2,...,nk=0,1,...m& \left(4^{\prime }\right)\end{array}$
where the decision variables of the model are the following integer variables:

[0158] $I_{\mathrm{ik}}=\{\begin{array}{cc}1& \mathrm{if}\mathrm{cell}i\mathrm{is}\mathrm{assigned}\mathrm{to}\mathrm{core}\mathrm{area}k\\ 0& \mathrm{otherwise}\end{array}$
In the mathematical model, Q_i is the maximum workday duration for driver i and α is the threshold probability that the total workload in each core area is less than the driver's maximum workday duration.

[0159] The functions g_ik and h_ik, are extremely complicated and may be too difficult to write for non-trivial problems. In one aspect, the method of the present invention uses linear approximations of g_ik and h_ik:
g_ik=ρ_ikT_iξ_i
h_ik=ρ_ikC_ik
where: ρ_ik is the learning curve factor of driver k in cell i;

[0160] T_i is the standard average time needed to serve a single stop in cell i;

[0161] μ_i is the mean demand in cell i;

[0162] C_ik is the assignment cost of assigning cell i to core area k (in terms of cell-to-cell travel time in an optimal cell tour).

[0163] The value of consistency or familiarity is reflected in the parameter ρ_ik. If the driver is more familiar with a cell, less time is needed to finish the work and less time is needed to travel from cell to cell along the route.

[0164] Unassigned cells are assigned to a dummy core area zero; in other words, k=0. For a cell i that is assigned to a core area k, k is non-zero: k≠0.

[0165] For assigned cells, when k≠0, the learning factor ρ_ik is a pre-defined value equal to the learning limit of driver k in cell i. It may be assumed that the learning limit is identical for all assigned cells.

[0166] The cell-to-cell travel time part of assignment cost is estimated as the follow:
C_ik=d_ik+d_i0−d_ok
where: d_ik=Expected travel time from cell i to the seed point of core area k;

[0167] d_i0=Expected travel time from cell i to the depot; and

[0168] d_0k=Expected travel time from depot to seed point of core area k.

[0169] For unassigned cells (i.e., when k=0) the value of route flexibility, in one aspect of the present inventive system, may be measured using the concept of visiting frequency. A cell that is not assigned nevertheless must be assigned on a temporary daily basis when a delivery to a customer in that area is required. For example, a delivery to an unassigned cell i may be temporarily assigned to core area k. The probability or frequency of this type of daily cell assignment may be referred to as F_ik. Such a delivery has an expected driver learning curve factor, w(F_ik). Accordingly, the expected learning curve factor for this delivery to cell i may be expressed as:

[0170] ${\rho }_{i0}=\sum _1^m{F}_{\mathrm{ik}}w\left(F_{\mathrm{ik}}\right).$

[0171] This equation demonstrates that, if a cell is split by several drivers in a daily route plan, the expected learning curve factor will be greater than that for a cell that is assigned to a core area on an ongoing basis. Accordingly, there is less learning benefit from unassigned cells in the core area building process.

[0172] However, unassigned cells in the daily dispatching stage provide more freedom for the dispatch system to balance the workload and optimally construct daily routes. In one embodiment, the present invention provides a method of indirectly estimating the value of this route flexibility in the following expression for calculating the expected cell-to-cell travel time contribution C_i0 of cell i in the optimal delivery route or cell tour:

[0173] $C_{i0}=\left(\frac{d_{{\mathrm{ib}}_2}-d_{{\mathrm{ib}}_1}}{d_{{\mathrm{ib}}_2}}\right)·\left(F_{{\mathrm{ib}}_1}{C}_{{\mathrm{ib}}_1}+F_{{\mathrm{ib}}_2}{C}_{{\mathrm{ib}}_2}\right)+\frac{d_{i0}}{{\stackrel{\_}{d}}_{b0}}{F}_{i0}·d_{i0}$
where real core areas are ranked by the increasing distance from the cell i and coordinates (b₁, b₂) represent the index of the core area with the best rank and the second-best rank. By simplifying the possible assignment of cells, in which a cell may be assigned only to core area (b₁, b₂) and core area zero (i.e., assigned to other core areas),

[0174] ${\stackrel{\_}{d}}_{b0}=\frac{d_{b_{1}0}+d_{b_{2}0}}{2}.$

[0175] F_ib1, F_ib2 and F_i0 are assigned value according to the weights of F_ib1, F_ib2 and d_i0. This function allows flexibility in two dimensions: (1) toward the depot; and (2) toward parallel drivers. The closer to the depot and the more equally positioned between two adjacent drivers, the more flexibility a cell may have. In the definition of C_i0,

[0176] $\left(\frac{d_{{\mathrm{ib}}_2}-d_{{\mathrm{ib}}_1}}{d_{{\mathrm{ib}}_2}}\right)\mathrm{and}\frac{d_{i0}}{{\stackrel{\_}{d}}_{b0}}$
serve as the flexibility coefficients for the two dimensions, respectively. Cells located a nearly-equal distance from the two closest core areas and cells near the hub or depot will have a smaller value of C_i0, which demonstrates those cells offer a higher value in terms of route flexibility. Accordingly, the function described provides an indirect approximation of the value of route flexibility.

[0177] For consistency, the expression for the learning curve factor ρ_i0 may be modified so that:
ρ_i0=F_ib1w(F_ib1)+F_ib2w(F_ib2)+F_i0w(F_i0)
which demonstrates mathematically that a cell assigned to a core area yields a measurable benefit in terms of the learning curve factor ρ_ik. Conversely, as shown above in the expression for the cell-to-cell travel time contribution C_i0, an unassigned cell yields a measurable benefit in terms of route flexibility.
7. Core Area Route Design Models

[0178] The methods and systems disclosed herein provide the tools for constructing a daily route plan for a service territory based on historical route data and the current daily demand. In one embodiment, as shown in FIG. 8, each cell 40 in a service territory 20 may be classified as a core cell 60, a flex zone cell 50, or a daily cell 70. Core cells 60 may be grouped together into a core area 100. The concept of using core areas 100 in a route planning process is an element of the present invention.

[0179] To construct the most efficient and effective core areas 100 for route planning purposes, in embodiments of the present invention, several models were developed for completing and optimizing a Core Area Route Design. Among the models disclosed herein are the Visit Frequency Limit model and the Stochastic Programming model.

[0180] 7.1. The Visit Frequency Limit Model

[0181] Referring again to the grid segment visiting frequency 154 for a first driver 23 as depicted in FIG. 5A, recall that a grid consistency index for each grid segment 152 may be calculated by averaging all the grid segment visiting frequencies 154 of all the drivers. The Visit Frequency Limit model of the present invention may use the grid consistency index to identify those geographical areas that are most often serviced by a consistent driver and, thus, most suitable for classification as a core cell 60, as shown in FIG. 6. Additional cell classifications are shown in FIG. 7, including flex zone cells 50 and daily cells 70.

[0182] Referring now to a route planning system 10 shown in FIG. 8, the present invention in another aspect includes a method of designing the core areas 100 which is referred to as the Visit Frequency Limit model. The following steps illustrate an embodiment of the Visit Frequency Limit model as shown in FIG. 22:

• • [0183] (1) A first step may include the execution of a route planning algorithm over a number of days during a reference period on a known data set. As in the Grid Method described above, a database may be used to categorize and easily reference the daily data for each driver.
  • [0184] (2) Run the Grid Method for measuring route consistency, as described above. The results may include a grid segment visiting frequency 154 for each driver 23, a grid consistency index for each grid segment 152 in the service territory 20 or portion thereof, and the identity of the most frequent driver for each grid segment, as shown in step 2202.
  • [0185] (3) Establish a minimum Visit Frequency Limit for the service territory, to use as the basis for deciding whether a grid segment may be classified as a core cell or placed in a core area, as shown in step 2204.
  • [0186] (4) As shown in step 2210, assign the most frequent driver to service the entire grid segment 152 on an ongoing basis if the grid segment visiting frequency 154 for the most frequent driver is greater than the minimum Visit Frequency Limit, as determined in step 2208. A grid segment 152 may include one or more cells 40. When a grid segment 152 is assigned to a most frequent driver, the cells 40 therein may be classified as core cells 60. In this aspect, the method of assigning grid segments also accomplishes the task of classifying cells as core cells.
  • [0187] (5) If the criterion of step (4) is not met, the grid segment 152 and the cells 40 therein remain unassigned, as shown in step 2212.

[0188] FIG. 9 is an illustration of a group of core areas 100 containing core cells 60 and/or grid segments 152 assigned to a most frequent driver using the Visit Frequency Limit model. FIG. 9 also shows a set of first exemplary delivery areas 80 within a service territory 20. Cells assigned on an ongoing basis to a core area 100 are referred to as core cells 60. Typically, one core area 100 is assigned to each available driver.

[0189] When service is requested by a customer located in an unassigned cell, the cell must be assigned on a daily basis to an available driver. Cells assigned on a daily basis are referred to as daily cells 70. As shown in FIG. 9, the first exemplary delivery areas 80 include a number of daily cells 70. At the operational level, daily cells 70 are generally assigned to the driver who serves the nearest core area 100. In one embodiment of the present invention, daily cells 70 may be optimally assigned to one extra driver by running a modified route planning algorithm specifically tailored to assign daily cells 70.

[0190] In one embodiment, a method of the present invention includes establishing a proximity factor relating each daily cell 70 to each core area 100. The proximity factor may include a distance element, a time element, or any additional element relevant to selecting the best core area 100 to accept the daily cell 70.

[0191] An individual stop 42 located outside a core cell 100, a flex zone cell 50, or a daily cell 70 may be referred to as a daily stop. In one embodiment, a method of the present invention includes establishing a stop proximity factor relating each daily stop to each core area 100. The stop proximity factor may include a distance element, a time element, or any additional element relevant to selecting the best core area 100 to accept the daily stop location.

[0192] In a related aspect of the invention, the Visit Frequency Limit model may be applied to existing route plans, such as cluster-based or loop-based plans, to preserve any route consistency already achieved within an active service territory. More specifically, the model may be applied to an existing loop route to evaluate its consistency and efficiency.

[0193] It should be understood that the design methodologies disclosed herein are scalable. The methods and systems of the present invention may be applied on a nationwide basis to evaluate routes across large clusters within a national service territory, on a neighborhood basis to evaluate route plans through particular streets within a portion of a cell, or on a stop-by-stop basis to evaluate parcel handling during a stop. Furthermore, the types of segmented areas and grids disclosed herein are used for descriptive purposes, as a way of illustrating the applications, and should not be construed as a limitation on the applicability or suitability for use on different shapes. In fact, it will be appreciated by those skilled in the art that the formulae and analyses disclosed herein may be mathematically and conceptually applied to any variety of discrete shapes and sizes.

[0194] 7.2. The Stochastic Programming Model

[0195] In another embodiment of the present invention, a mathematical model for constructing optimal core areas for route planning purposes. In the Stochastic Programming model, a set of optimal core areas is identified by solving a Non-linear Generalized Assignment Program (NGAP) with the objective of maximizing the savings produced by driver familiarity and minimizing the assignment cost associated with assigning a cell to a core area.

[0196] By identifying core areas at a strategic level, a good framework of driver service territories is established, which is essential for making consistent route plans while at the same time leaving enough flexibility to accommodate the varying customer demand, which in turn results in better driver utilization. In one embodiment, an effective meta-heuristic (tabu search) method was developed for clustering cells into core areas.

[0197] One of the constraints in this model is that the probability that the total workload in the core area may exceed the workday duration of the driver. Workday duration will be used as a threshold to analyze how the size of a core area may affect the total routing cost.

[0198] In one embodiment, a stochastic Core Area Design model is provided which considers the stochastic customer locations and demands. By optimally partitioning a service territory, the stochastic model maximizes the benefits from driver familiarity by producing more consistency in the daily operational vehicle dispatching plan, while at the same time allowing enough flexibility to produce optimal or near-optimal daily routes.

[0199] If a cell is assigned to a core area, the total travel time a driver needs to serve all the stops within this cell will decrease to the proportion ρ=T_∞(0<ρ<1), which may be referred to as a learning curve factor ρ.

[0200] T_∞ is the driver learning limit in terms of the percentage of standard average time used to serve a stop within a cell or travel from cell to cell. Obviously, it is advantageous to assign as many cells to the core area as feasible. However, there are two limiting factors:

[0201] The workload for a route should rarely exceed shift duration. This limits the threshold probability that the workload within a core area exceeds the total work duration (say, within 5%). The larger this threshold, the more likely that the core area will generate a route with overage in terms of total route duration.

[0202] In one embodiment, the method of the present invention includes a primal stochastic program formulated to accomplish the stochastic Core Area Design model. The decision variables of the model are the following integer variables:

[0203] $I_{\mathrm{ik}}=\{\begin{array}{cc}1& \mathrm{if}\mathrm{cell}i\mathrm{is}\mathrm{assigned}\mathrm{to}\mathrm{core}\mathrm{area}k\\ 0& \mathrm{otherwise}\end{array}$
Parameters of the model are as follows:

• [0204] C_ik=The assignment cost of assigning cell i to core area k
• [0205] ρ_ik=Learning curve factor of cell i for driver k
• [0206] T_i=Standard average time needed to serve a single stop in cell i
• [0207] Q_i=Maximum working duration for driver i
• [0208] μ_i=Mean demand in cell i
• [0209] δ_i=Standard deviation of demand in cell i
• [0210] ξ_i=Numbers of Customer in cell i (random variable)
• [0211] α=Threshold which is the probability that the total workload in each core area is less than maximum working duration.
  The stochastic programming model can be formulated as follows:

[0212] $\begin{array}{cc}\mathrm{Min}E\left[\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\right]& \left(1\right)\\ S.t.\sum _k{I}_{\mathrm{ik}}=1,i=1,2,...,nk=0,1,...m& \left(2\right)\\ P\left(\sum _i\left({\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}+{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}\right)I_{\mathrm{ik}}\le Q_k\right)\ge 1-\alpha ,\left(k=1,...,m\right)& \left(3\right)\\ I_{\mathrm{ik}}=0\mathrm{or}1,i=1,2,...,nk=0,1,...m& \left(4\right)\end{array}$

[0213] The objective function in Equation (1) represents the sum of the costs incurred from by assigning a cell to a core area (including a dummy core area zero). ρ_ik T_i ξ_i in the objective function represents the learning-adjusted workload within the cell (including stop to stop traveling time and service time at each stop) by assigning it to core area due to driver familiarity with their service territories. ρ_ik C_ik represents the contribution of learning-adjusted cell to cell traveling time in the optimal cell tour based on the cell assignment.

[0214] Equation (2) ensures that each cell is assigned to only one core area.

[0215] Equation (3) contains the probability constraints for working duration of each core area. It states that the probability that the total workload in each core area exceeds the maximum working duration of the driver can not be higher than α.

[0216] 7.3. Flex Zones

[0217] In one embodiment, the Core Area Design models of the present invention include a method of constructing a flex zone 200, which is a flexible service area generally located around the depot or hub 30. This zone has the most flexibility because work can be assigned to alternate drivers without incurring a large change in total routing cost. The size of this flex zone is decided by the variability of the customer demand. FIG. 8 illustrates a flex zone 200, which may contain one or more flex zone cells 50.

[0218] The establishment of a flex zone 200 effectively removes the flex zone cells 50 from the calculation process which, in turn, simplifies and speeds the process of building core areas using the models of the present invention. The flex zone 200 may take any shape and may vary over time, as customer demand changes. In some plans, the hub 30 may be located outside the flex zone 200 altogether.

[0219] In one embodiment, the present invention includes a method for constructing a flex zone 200 by using a time limit; in other words, by including all the customers who can be accessed within a certain number of minutes. Generally, any of a variety of limits or constraints particularly suited to the unique characteristics of a service territory 20 may used to define a flex zone 200. Those skilled in the art of route planning will appreciate the various characteristics to be considered, which include but are not limited to terrain, travel time, travel distance (proximity), access along major roads and corridors, travel barriers such as railroad grade crossings and draw bridges, natural barriers such as rivers and mountains, as well as any other local factor affecting the dispatching and routing of vehicles.

8. Solution Method for the Stochastic Programming Model

[0220] The Visit Frequency Limit model may solved using empirical data, such as the grid segment visiting frequencies. For the Stochastic Programming model, in one embodiment of the present invention, a mathematical solution is provided. In the Stochastic Programming model, a set of optimal core areas is identified by solving a Non-linear Generalized Assignment Program (NGAP).

[0221] In this embodiment, it may be assumed that the demands ξ_i are independent, normal, distributed random variables with means μ_i and standard deviations δ_i. M_k and S_k are the mean and standard deviation, respectively, of the demand on core area k, where:

[0222] $M_k=\sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}$ $S_k=\sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}\right)}^2}$
When the demands ξ_i are normally distributed, there exists a constant τ, such that:

[0223] $\mathrm{Pr}\left[\left(\sum _i{\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}-M_k\right)/S_k\le \tau \right]=1-\alpha .$
Therefore, the chance constraint becomes:

[0224] $\sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}\le Q_k-\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}$
and the chance-constrained program becomes a non-linear integer program:

[0225] $\begin{array}{cc}\mathrm{Min}E\left[\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\right]& \left(5\right)\\ S.t.\sum _k{I}_{\mathrm{ik}}=1,i=1,2,...,mk=0,1,...m& \left(6\right)\\ \sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}+\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}\le Q_k,(k=1,...m)& \left(7\right)\\ I_{\mathrm{ik}}=0\mathrm{or}1,i=1,2,...,nk=0,1,...m& \left(8\right)\end{array}$
Notice that:

[0226] $\begin{array}{c}E\left[\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\right]=\sum _{i,k}E\left[{\rho }_{\mathrm{ik}}{T}_i{\xi }_i{I}_{\mathrm{ik}}\right]+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\\ =\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_{i}E\left[{\xi }_i\right]I_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\\ =\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}\end{array}$
So, the above model becomes:

[0227] $\begin{array}{cc}\mathrm{Min}\sum _{i,k}{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\sum _{i,k}{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}{I}_{\mathrm{ik}}& \left(9\right)\\ S.t.\sum _k{I}_{\mathrm{ik}}=1,i=1,2,...,nk=0,1,...m& \left(10\right)\\ \sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}+\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}\le Q_k,(k=1,...,m)& \left(11\right)\\ I_{\mathrm{ik}}=0\mathrm{or}1,i=1,2,...,nk=0,1,...m& \left(12\right)\end{array}$

[0228] The above set of equations represents a type of Non-linear Generalized Assignment Problem (Non-linear GAP). There is no effective algorithm in the field for solving this kind of problem. In one aspect of the present invention, two alternative approaches to finding a solution were considered:

• • [0229] 1. Find a linear approximation for the non-linear constraint expressed in Equation (11), and then transform the problem into an integer linear program, which can be solved using a sophisticated algorithm.
  • [0230] 2. Construct a heuristic algorithm for a stochastic Core Area Design model which incorporates the spatial characteristics of the problem. At each iteration of the heuristic, the objective function in Equation (9) and the non-linear constraint in Equation (11) may be considered to guide the next search, such that an optimal solution can be found.

[0231] In one embodiment, the system of the present invention combines these two approaches in the following way. First, a meta-heuristic algorithm (applying a tabu search) is developed for solving the Non-linear GAP. Then, the non-linear constraint in Equation (11) is “linearized” to form a Linear GAP, the solution of which serves as a lower bound. The lower bound is then used to evaluate the quality of the solution produced by the meta-heuristic algorithm.

[0232] 8.1. Combinatorial Optimization and Meta-Heuristics

[0233] In general, a Generalized Assignment Problem (GAP) involves assigning tasks to agents. Each task must be assigned to one and only one agent. Each agent has a limited amount of capacity. An agent may have multiple tasks assigned to it, but the sum of the resource requirements for these tasks must not exceed the agent's capacity. The resource requirements of a particular task and the assignment cost depend on the agent who will perform the task. The objective of this combinatorial optimization problem is to minimize the total assignment cost. The Non-linear GAP differs from the Linear GAP in that the objective function and the capacity constraint can be non-linear. In the context of route planning, the capacity constraint within each core area is non-linear.

[0234] The Non-linear GAP formulation for the stochastic Core Area Design model contains a large number of binary variables as well as non-linear constraints, making the task of obtaining an optimal solution more difficult. In one embodiment of the present invention, the following meta-heuristic algorithm with tabu search is provided as an approach to solving large-scale Non-linear GAPs. The method is based upon a tabu search methodology, which incorporates a number of sophisticated features, including long-term memory, systematic dynamic tabu tenure, and a highly-effective move decision mechanism derived from the notion of ejection chains.

[0235] Combinatorial optimization is the mathematical study of finding an optimal arrangement, grouping, ordering, or selection of a finite number of discrete objects. Many combinatorial problems (including Linear GAP and Non-linear GAP) are hard to solve (i.e., intractable). An optimal algorithm for intractable combinatorial problems would require an exponentially-growing number of computational steps as the problem size increases. Accordingly, only small problems can be solved within a reasonable time. In some applications, heuristics (or approximate algorithms) are used to find an approximate solution to a combinatorial problem.

[0236] Meta-heuristics are a class of algorithms that may be used to attack intractable combinatorial optimization problems where classical heuristics are not efficient. A meta-heuristic typically involves an iterative process using rules and information structures designed to produce an efficient and near-optimal solution. The methodologies underlying meta-heuristics may be derived from concepts such as classical heuristics, artificial intelligence, biological evolution, neural systems, and statistical mechanics. This family of approaches include genetic algorithms, greedy random adaptive search procedures, problem-space search, neural networks, simulated annealing, tabu search methods, and their hybrids.

[0237] 8.2. Tabu Search Heuristics

[0238] Tabu or taboo may be defined as an adjective describing something that is set apart, forbidden, or banned because it constitutes a risk. In the mathematical context, a tabu search includes restrictions that forbid the selection of possible solutions known to be bad or inefficient. By incorporating the advantages of adaptive memory and responsive exploration, a tabu search emphasizes both intensification and diversification in the search strategy.

[0239] A tabu search typically starts from an initial solution, which may or may not be feasible. For each iteration, the tabu search rules are applied to evaluate possible “moves” within a matrix or system of possible solutions. Potential moves are found by exploring the “neighborhood” around the current solution. During each iteration, the potential moves are evaluated in terms of which is the most promising for improving the current solution, and that “best move” is selected. A move classified as tabu is excluded from the evaluation process. Any of the recently-generated potential moves may be classified as tabu. A local optimum may be classified as tabu, which allows this search method to drive the search beyond the local optimum and toward a global optimum.

[0240] 8.2.1. Network Formulation

[0241] The mathematical model of the Non-linear GAP may be graphically represented by means of network-related formulation. The field of so-called “netforms” has allowed researchers to formulate problems as networks. There are also notational conventions that make possible the representation of problems in a general network-related framework. These include flow restrictions such as integer restrictions. The use of these conventions provides a concise representation of network programming problems with side constraints.

[0242] The Non-linear GAP netform consists of n+m+1 nodes and n×(m+1) arcs. For each source node j there is a fixed supply of one unit, and for each destination node k there is a demand that must not exceed Q_k units. The flow in each arc is restricted to integers, so they can be only 0 or 1. There is also a cost f_ik and a multiplier α_ik for every arc, which respectively represent the cost and capacity requirement for assigning cell i to core area k. FIG. 17 shows the Non-linear GAP netform, where the costs are enclosed in boxes and the multipliers are enclosed in triangles.

[0243] From the Non-linear GAP formulation above:
f_ik=ρ_ik(T_iμ_i+C_ik) and α_ik=ρ_ik(T_iμ_i+C_ik)+φ_i
where φ_i represents the contribution of cell i to the non-linear part:

[0244] $\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2)}$
in the constraint Equation (11).

[0245] 8.2.2. Neighborhood Structure

[0246] In a tabu search algorithm, a move is defined as a modification to a current solution according to a set of pre-defined procedures or rules that produce another, new solution. The collection of solutions that result from all possible moves around the current solution is called a neighborhood of the current solution.

[0247] Based upon a current solution of a Non-linear GAP, a basic move is to simply change the assignment of cell i from core area j to core area k. In terms of a netform, this simple move is equivalent to deleting the arc extending from node i to node j, and inserting an arc that extends from node i to node k. After every move, the assignment constraints in Equation (10) must be satisfied, but the capacity constraints in Equation (11) cannot be guaranteed to hold. In another words, infeasibility may occur if the capacity constraints are violated. This special feature of a move enables the tabu search to cross the feasible region of local optimal solutions and enter the realm of possible global optimal solutions.

[0248] A compound move refers to a combination or a series of simple moves, in which the number of components is decided by a certain compound move construction termination procedure. Compound moves do not concern isolated choices of moves with particular features, but rather involve coordinated choices of moves with interlinking properties. Procedures that incorporate compound moves are often called variable-depth methods, because the number of components involved in a compound move generally vary from step to step. A move neighborhood constructed from compound moves is more likely to produce effective and efficient moves.

[0249] The ejection chain method is a special subclass of the variable-depth methods. FIG. 18 shows an example ejection chain for the Non-linear GAP problem. An ejection chain is started by selecting a set of elements, whose states will undergo some changes (for example, to new values or new positions). The consequences of this change will cause at least one element in the set to be ejected from its current state.

[0250] In FIG. 18, the dashed lines represent the original assignment: cell i₁ to core area j₁, cell i₂ to core area j₂, and so forth. A simple move here is to reassign cell i₁ to core area j₂, as shown by the solid line. An ejection chain results because the move “ejects” an element: for the arc (i₂, j₂) at node j₂, the connection to node i₂ must be re-assigned to a new core area (j₃, for example). This process may continue through additional nodes until a suitable termination criterion is met.

[0251] A simple termination criterion for the Non-linear GAP may be continuing the chain until no ejection occurs. In the context of core area assignments, there may exist a core area further down the chain which may have enough capacity to serve the last ejected node. At the least, the dummy core area zero has the capacity because it has no capacity constraints.

[0252] The capacity slack s_k of core area k may be expressed as:

[0253] $s_k=Q_k-\sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}-\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}-\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}$
A solution is feasible only if all s_k≧0. The degree of infeasibility, ν, of an assignment-feasible solution x, may be expressed as the absolute value of the sum of all the negative capacity slacks s_k:

[0254] $v\left(x\right)=\mathrm{abs}\left(\sum _{k=1}^m\min \left(s_k,0\right)\right)$
Therefore, solution x is feasible only if ν(x)=0. The feasibility measure ν(x) is a component in the tabu search procedure.

[0255] In one embodiment, the present invention applies the following types of ejection chains in the tabu search procedure. FIG. 19 is a graphical representation of several types of ejection chains that may be used in the method of the present invention.

[0256] For a single-ejection chain, a leaving arc (i, j₁), and an entering arc (i, j₂), a feasibility condition may be expressed as s_j2−(ρ_ij2T_iμ_i+ρ_ij2+C_ij2+Φ_ij2)≧0.

[0257] For a first double-ejection chain, having a leaving arc (i₁, j₁) and an entering arc (i_i, j₂), a feasibility condition may be expressed as s_j2+(ρ_i2j2T_i2μ_i2+ρ_i2j2C_i2j2+δ_i2j2)−(ρ_i1j2T_i1μ_i1+ρ_i1j2C_i1j2+Φ_i1j2)≧0.

[0258] For a second double-ejection chain, having a leaving arc (i₂, j₂) and an entering arc (i₂, j₃), a feasibility condition may be expressed as s_j3−(ρ_i2j3T_i3μ_i3+ρ_i2j3C_i2j3+Φ_i2j3)≧0.

[0259] For a first circular-ejection chain, having a leaving arc (i₁, j₁) and an entering arc (i₁, j₂), a feasibility condition may be expressed as s_j1+(ρ_i1j1T_i1μ_i1+ρ_i1j1C_i1j1+δ_i1j1)−(ρ_i2j1T_i2μ_i2+ρ_i2j1C_i2j1+Φ_i2j1)≧0.

[0260] For a second circular-ejection chain, having a leaving arc (i₂, j₂) and an entering arc (i₂, j₁), a feasibility condition may be expressed as s_j2+(ρ_i2j2T_i2μ_i2+ρ_i2j2C_i2j2+δ_i2j2)−(ρ_i1j2T_i1μ_i1+ρ_i1j2C_i1j2+Φ_i1j2)≧0

[0261] 8.2.3. Tabu Search Procedure

[0262] Initialization. In the initialization step, a starting solution (it may be feasible or not) may be generated and the objective function value may be calculated. Also, all data structures, including a tabu matrix, a long-term memory, and a solution matrix may be initialized. The tabu data structure may take the form of a time stamp that indicates the iteration number at which a move “clears itself” from having a tabu status. Specifically, a the two-dimensional array such as one referred to as “tabu_size” may be created, where the (i, j) element of this array holds the iteration number at which arc (i, j) will clear its tabu status, and it can re-join the neighborhood construction procedure for selecting the best move. At the beginning of the search, “tabu_size” must be initialized to zero.

[0263] In one embodiment of the present invention, the method uses a starting solution in which all cells are assigned to the dummy core area zero. In other words, at the beginning, all cells are unassigned. Therefore, the objective of the starting solution may be expressed as:

[0264] $\sum _{i=1}^n\left({\rho }_{i0}{T}_i{\mu }_i+{\rho }_{i0}{C}_{i0}\right)$

[0265] Best Move. Tabu search methods are designed to select the best move available at each step, based upon the current search state. There are many criteria for selecting the best move. One simple example is to select the move that has the best improvement in objective function, Z(x). In one embodiment, the method of the present invention combines the best improvement function Z(x) with the feasibility measure ν(x) in the following manner:

[0266] Where x is the current solution, x′ is the neighboring solution for which Z(x′)−Z(x) is minimized, and x′ is the neighboring solution for which ν(x′)−ν(x) is minimized (in other words, the ties are broken by comparing the change in the objective function value), the best move is defined as follows, for the following solution states:

[0267] For the solution state defined by ν(x)>0, the best move is x″.

[0268] For the solution state defined by ν(x)=0 and ν(x″)>0, the best move is x′.

[0269] For the solution state defined by ν(x)=0 and ν(x″)=0 and Z(x″)≦θ, the best move is x″.

[0270] For the solution state defined by ν(x)=0 and ν(x″)=0 and Z(x″)>θ, the best move is x′.

[0271] When the current solution is not feasible, the best move is the one that reduces the infeasibility the most. When the current solution is within the feasible region but no feasible move exists, the search will select the move that has the smallest objective function value. Non-improving moves are accepted as long as the moves are within the feasible region and the associated objective function values do not exceed the threshold θ. However, when the associated objective function values exceed this threshold, the search selects as the best move the move that reduces the objective function value the most. This best move may possibly lead the search into the infeasible region, which makes it possible to explore other feasible regions.

[0272] Threshold θ is defined as the objective function value of the first solution found when the search last entered a feasible region. The objective function value Z(x_B) of the best solution x_B is updated each time the search enters the feasible region. If m×n iterations have passed without improving Z(x_B), then θ is reset to the objective function value of the current solution Z(x), which may prompt the non-improving search state to cross the capacity-feasibility boundary and enter the infeasible region.

[0273] Evaluating moves includes the task of evaluating the change in objective function value and the change in overall capacity slack. The evaluation of change in objective function value is straightforward:

For single ejection chain (i₁, j₁)(i₁, j₂):
Z(x′)−Z(x)=ρ_i1j2T_i1μ_i1+ρ_i1j2C_i1j2−ρ_i1j1T_i1μ_i1−ρ_i1j1C_i1j1
For double ejection chain (i₁, j₁)(i₁, j₂), (i₂, j₂)(i₂, j₃):
Z(x′)−Z(x)=ρ_i1j2T_i1μ_i1+ρ_i1j2C_i1j2+ρ_i2j3T_i2μ_i2+ρ_i2j3C_i2j3−ρ_i1j1T_i1μ_i1−ρ_i1j1C_i1j1−ρ_i2j2T_i2μ_i2−ρ_i2j2C_i2j2
For circular ejection chain (i₁, j₁)(i₁, j₂), (i₂, j₂)(i₂, j₁):
Z(x′)−Z(x)=ρ_i1j2T_i1μ_i1+ρ_i1j2C_i1j2+ρ_i2j1T_i2μ_i2+ρ_i2j1C_i2j1−ρ_i1j1T_i1μ_i1−ρ_i1j1C_i1j1−ρ_i2j2T_i2μ_i2−ρ_i2j2C_i2j2
The evaluation of change in overall capacity slack is more complicated. The evaluation for a single-ejection chain is illustrated using pseudo-code, as follows:

[0274]

	c1 ← 0;
	c2 ← 0;
	if (j1!=0)
	{
	ϕ 1 ← ∑ i ⁢ ⁢ assigned ⁢ ⁢ to ⁢ ⁢ j 1 ⁢ ⁢ ( ρ ij 1 ⁢ T i ⁢ δ i ) 2 ; ϕ 2 ← ϕ 1 - ( ρ i 1 ⁢ j 1 ⁢ T i 1 ⁢ δ i 1 ) 2 ; ϕ 1 ← τ ⁢ ϕ 1 ; ϕ 2 ← τ ⁢ ϕ 2 ; d 1 ← - min ⁡ ( s j 1 , 0 ) ; d 2 ← ρ i 1 ⁢ j 1 ⁢ T i 1 ⁢ μ i 1 + ρ i 1 ⁢ j 1 ⁢ C i 1 ⁢ j 1 + ϕ 1 - ϕ 2 ; c 1 ← min ⁡ ( d 1 , d 2 ) ;
	}
	if (j1!=0)
	}
	ϕ 1 ← ∑ i ⁢ ⁢ assigned ⁢ ⁢ to ⁢ ⁢ j 2 ⁢ ⁢ ( ρ ij 2 ⁢ T i ⁢ δ i ) 2 ; ϕ 2 ← ϕ 1 - ( ρ i 1 ⁢ j 2 ⁢ T i 1 ⁢ δ i 1 ) 2 ; ϕ 1 ← τ ⁢ ϕ 1 ; ϕ 2 ← τ ⁢ ϕ 2 ; d 1 ← max ⁡ ( s j 1 , 0 ) ; d 2 ← ρ i 1 ⁢ j 2 ⁢ T i 1 ⁢ μ i 1 + ρ i 1 ⁢ j 2 ⁢ C i 1 ⁢ j 2 + ϕ 2 - ϕ 1 ; c 2 ← min ⁡ ( d 1 , d 2 , 0 ) ;
	}
	v(x′) − v(x) ← −(c1 + c2);

[0275] The double-ejection chain and circular-ejection chain use a similar methodology and, hence, a similar pseudo-code.

[0276] In this tabu search meta-heuristic, the best move procedure is computationally more expensive than any other procedure (for example, the executing move procedure or the updating procedure). The best move selection procedure may consume a length of time that may be expressed as O(n²). Because the best move procedure is time-intensive, the method of the present invention improves to computational efficiency of the move evaluation procedures. In the pseudo-code above, most of the computation time is spent calculating the change in the non-linear portion of the capacity slack. For a larger problem, which may require thousands of moves to be evaluated during each step, the search speed may likely be slower.

[0277] To accommodate for slower search speeds, the method of the present invention in one embodiment includes a linear approximation for the non-linear

[0278] $\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}\approx \tau \left(\sum _i{\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)/\sqrt{n_k}$
portion in constraint Equation (11) above, as follows:
where n_k is the total number of cells assigned to core area k. In the context of the Non-linear GAP model:

[0279] $n_k=\sum _i{I}_{\mathrm{ik}}$

[0280] To test the quality of this approximation, the value of these two expressions may be recorded each time the evaluation procedure is executed during the tabu search procedure.

[0281] In one embodiment of the present invention, the method includes replacing the non-linear portion of the capacity slack with the linear approximation developed above for the move evaluation procedure. In use, the linear approximation has no significant affect on the tabu search solution quality. Using the linear approximation, the pseudo-code above becomes simplified:

[0282]

	c1 ← 0;
	c2 ← 0;
	if (j1 != 0)
	{
	d1 ← −min(s j1,0);
	d2 ← ρi1j1Ti1μi1 + ρi1j1Ci1j1 + τ(ρi1j1Ti1δi1 / {square root over (nj1)});
	c1 ← min(d1,d2);
	}
	if (j2 != 0)
	{
	d1 ← max(s j1,0);
	d2 ← ρi1j2Ti1μi1 + ρi1j2Ci1j2 + τ(ρi1j2Ti1δi1 / {square root over (nj2)});
	c2 ← min(d1 − d2,0);
	}
	v(x′) − v(x) ← −(c1 + c2);

[0283] A move can only be selected as a candidate for the best move if (1) the entering arc associated with the move is not in tabu status, or (2) a set of aspiration criteria are met so as to clear its tabu status. In the tabu search procedure, aspiration criteria are used if the move currently under evaluation leads the search to the best feasible solution ever found. In this case, the tabu status of the move is overridden.

[0284] Executing a Move and Updating Tabu. The execution of the best move will update the current trial solution. In this embodiment, the items that need to be updated are:

• • [0285] The current solution structure.
  • [0286] The capacity slack structure for all core areas.
  • [0287] The best move selection objective function value threshold θ.
  • [0288] The best objective function value found so far.
  • [0289] The data structure that holds the number of cells that currently assigned to each core area.
  • [0290] The tabu structure and the long term memory.
    All these items are directly affected by changes in a current trial solution. Updating the tabu structure and the long-term memory may be accomplished as follows:

[0291] After a move has been executed, the leaving arc (for example, the arc (i₁, j₁) in the single-ejection chain) becomes tabu. During the tabu procedure, the leaving arc is not allowed to be part of the solution. The tabu tenure (which is the number of iterations during which the leaving arc remains tabu) is a function of three elements:

• • [0292] (1) The number of possible arcs leaving the same source node (cell) where the leaving arc originates (in the Non-linear GAP, the number of such possible arcs is m+1);
  • [0293] (2) The difference between the ranks of the leaving arc and the entering arc; and
  • [0294] (3) The frequency with which the leaving arc has been a member of a previously-selected ejection chain.

[0295] Suppose the leaving and entering arcs are (i_L, j_L) and (i_E, j_E), respectively. Then, the tabu tenure of leaving arc (i_L, j_L) may be expressed as:

[0296] $\mathrm{tabu\_time}\left(i_L,j_L\right)=\left(m+1\right)\left(\frac{3}{2}+\frac{\Delta }{2m}\right)+m\frac{\Omega \left(i_L,j_L\right)}{{\Omega }_{\max }}$
where: Ω(i_L, j_L)=the number of times arc (i_L, j_L) has been part of an executed move;

[0297] Ω_max=the maximum Ω(i_L, j_L) for all (i, j);

[0298] Δ=r(i_L, j_L)−r(i_E, j_E); and

[0299] r(i, j)=the position of arc (i, j) when all arcs leaving node j are ordered by increasing value of the expected cell-to-cell travel time contribution of assigning cell i to core area j in an optimal route or cell tour (that is, the parameter C_ij in the Non-linear GAP model).

[0300] From the expression above, the minimum tabu time value is (m+1) when:

• • [0301] (1) the leaving arc has the best rank (i.e., when r(i_L, j_L)=1);
  • [0302] (2) the entering arc has the worst rank (i.e., when r(i_E, j_E)=m+1); and
  • [0303] (3) the leaving arc has never been part of an executed ejection chain (i.e., when Ω(i_L, j_L)=0).
    On the other hand, the maximum tabu time value is (3 m+1), when:
    r(i_L,j_L)=m+1,r(i_E,j_E)=1 and Ω(i_L,j_L)=Ω_max.

[0304] Overall, by implementing this tabu search procedure, with special features such as a systematic dynamic tabu list, long-term memory, and ejection chains, the method of present invention represents an effective and computationally-efficient heuristic method for solving the Non-linear GAP model for identifying the core areas.

[0305] 8.3. Solution Quality

[0306] The linearization of the non-linear constraint disclosed herein transforms the Non-linear GAP into a Linear GAP, the solution for which may serve as a lower bound for the Non-linear GAP. Comparing the solution of the tabu search heuristic with this lower bound provides a measure of the solution quality of the tabu search heuristic.

[0307] The non-linear constraints for the Non-linear GAP are expressed in Equation (11):

[0308] $\sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\tau \sqrt{\sum _i{\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}^2}+\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}\le Q_k,\left(k=1,\dots ,m\right)$
The following equations represent the linearized counterparts of the non-linear constraints in Equation (11):

[0309] $\begin{array}{cc}\sum _i{\rho }_{\mathrm{ik}}{T}_i{\mu }_i{I}_{\mathrm{ik}}+\frac{\tau \sum _i\left({\rho }_{\mathrm{ik}}{T}_i{\delta }_i{I}_{\mathrm{ik}}\right)}{\sqrt{n}}+\sum _i{\rho }_{\mathrm{ik}}{C}_{\mathrm{ik}}\le Q_k,\left(k=1,\dots ,m\right)& \left(13\right)\end{array}$
Notice that if equations (11) are met, then equations (13) are also met. This means that the feasible region restricted by constraints (13) by definition contains the feasible region restricted by constraints (11). Therefore, the Linear GAP expressed in Equations (9), (10), (13), and (12) must contain the optimal solution of the Non-linear GAP expressed in Equations (9), (10), (11), and (12). Hence, the optimal solution of the approximation GAP may be used to serve as the lower bound of the Non-linear NGAP.

[0310] Software such as CPLEX Mixed Integer Programming (MIP) solver for solving the approximation GAP equations for a lower bound. In some cases, where a software system cannot find an optimal solution within a node limit, the Linear Programming (LP) solution of the problem may be used as a more-restrictive lower bound. Comparing these two lower bounds provides a reasonable measurement of the tabu search solution quality.

9. Operational Cell Routing

[0311] Operation cell routing involves the assignment of non-core or daily cells 70 to a route, and also involves the ordering or sequencing of stops in a route. FIG. 9 depicts a flex zone 200 containing flex zone cells 50, several core areas 100 containing core cells 60, and a set of first exemplary delivery areas 80 within a service territory 20. The first exemplary delivery areas 80 include a number of daily cells 70.

[0312] FIG. 10 shows a group of second exemplary delivery areas 90 for a particular day within the service territory 20. As shown, the second exemplary delivery areas 90 include different daily cells 70 than those assigned to the first exemplary delivery areas 80 in FIG. 9. The system 10 of the present invention accommodates a variety of constraints and changing conditions when assigning the daily cells 70 to a particular core area 100.

[0313] In one embodiment, the present invention includes a method of selecting the core area 100 best suited to include a particular daily cell 70. In one embodiment, a method of the present invention includes establishing a proximity factor relating each daily cell 70 to each core area 100. The proximity factor may include a distance element, a time element, or any additional element relevant to selecting the best core area 100 to accept the daily cell 70.

[0314] This operational cell routing problem may be formulated mathematically as an integer linear program with a three-index vehicle flow. In this formulations, a graph G may include a set of vertices V (representing cells) and a set of arcs A. The hub 30 in the service territory may be located at vertex 1. In mathematical terms, let G=(V, A) be a graph where V={1, . . . , n} is a set of vertices and A is a set of arcs.

[0315] A non-negative travel time matrix C=(c_ij) is associated with every arc (i, j), i≠j. The term c_ij can be interpreted as the overall travel time from cell i to cell j, which includes the cell-to-cell travel time d_ij and the travel time w_i within cell i. Accordingly, c_ij=d_ij+w_i. In practice, the values of d_ij and w_i may depend on other cells.

[0316] In one embodiment, the method of the present invention includes a travel time matrix constructed based upon approximation models disclosed above. The matrix may include a fleet of m available vehicles, where the value of m may be fixed or variable. When m is variable, the cost associated with adding a vehicle may be ignored. All vehicles may be assumed to be identical and driven by drivers having the same total workday duration, such that the duration of any route many not exceed a prescribed limit Q.

[0317] A set U_k may be defined that includes all core cells 60 assigned to driver k. The problem is to design a set of least-time duration vehicle routes such that:

• • [0318] (1) each core cell 60 is visited exactly once, by exactly one vehicle (and its one driver);
  • [0319] (2) each route starts and ends at the hub 30;
  • [0320] (3) no route exceeds time duration limit Q; and
  • [0321] (4) all core cells 60 assigned to a particular vehicle are served only by that vehicle.

[0322] Using a three-index formulation, the variables x_ijk indicate whether location (i, j) is traversed by vehicle k or not. The binary variables y_ik are set to equal 1 if a cell i is served by vehicle k. Accordingly, the three-index formulation may be expressed as follows:

[0323] $\begin{array}{cc}\min .\stackrel{m}{\sum _{k=1}}\sum _{i\ne j}{c}_{\mathrm{ij}}{x}_{\mathrm{ijk}}& \left(14\right)\\ s.t.\sum _{i\ne j}{c}_{\mathrm{ij}}{x}_{\mathrm{ijk}}\le Q\left(k=1,\dots ,m\right)& \left(15\right)\\ \sum _{i=1}^n{y}_{\mathrm{ik}}=\{\begin{array}{cc}m& \left(i=1\right)\\ 1& \left(i=2,\dots ,n\right)\end{array}& \left(16\right)\\ \sum _{i=1}^n{x}_{\mathrm{ijk}}=y_{\mathrm{jk}}\left(j=1,\dots ,n;k=1,\dots ,m\right)& \left(17\right)\\ \sum _{j=1}^n{x}_{\mathrm{ijk}}=y_{\mathrm{ik}}\left(i=1,\dots ,n;k=1,\dots ,m\right)& \left(18\right)\\ \sum _{i,j\in S}{x}_{\mathrm{ijk}}\le S-1\left(S⋐V;S\ge 2;k=1,\dots ,m\right)& \left(19\right)\\ y_{\mathrm{ik}}=1\left(i\in U_k,k=1,\dots ,m\right)& \left(20\right)\\ x_{\mathrm{ijk}}\in \left\{0,1\right\}\left(i,j=1,\dots ,n;k=1,\dots ,m\right)& \left(21\right)\\ y_{\mathrm{ik}}\in \left\{0,1\right\}\left(i=1,\dots ,n;k=1,\dots ,m\right)& \left(22\right)\end{array}$

[0324] The constraint in Equation (15) represents the total time restriction Q, for each driver's maximum workday duration. The constraint in Equation (19) is a sub-tour elimination constraint.

[0325] The operational cell routing model expressed in this three-index formulation is similar to a classical vehicle routing problem (VRP) with a maximum driver workday duration constraint, except for the pre-assignment constraint in Equation (20) and the following differences:

• • [0326] (1) The operational cell routing model is for routing cells instead of individual stops. Routing at the cell level is more complicated because the cost of an arc exiting a cell depends on the preceding cell as well as the next cell.
  • [0327] (2) The model must accommodate the fact that some of the cells are assigned as core cells 60 to a regular driver.
  • [0328] (3) The effect of driver learning is incorporated in the route planning process, so that certain cells are assigned to the most-familiar driver.
    In one embodiment, the method of the present invention addresses these differences by adding features to the classical VRP solution method, as disclosed below in section 8.2.

[0329] 9.1. The VRP Algorithm

[0330] In one embodiment, the method of the present invention begins with a set of VRP algorithms that combine a sequential route construction heuristic and a parallel route construction heuristic. The solutions may be improved by implementing an inter-route tabu search procedure.

[0331] A sequential route construction heuristic builds routes one at a time until all the stops are routed, whereas a parallel route construction heuristic builds multiple routes simultaneously. The basic methodologies, however, are the same in that both heuristics proceed by inserting one stop at a time into the emerging route or routes, and the choices of which stop to insert and where to insert it are based on heuristic cost measures.

[0332] For example, a sequential route construction heuristic may include the following steps:

• • [0333] (1) select a seed stop for a new route;
  • [0334] (2) for each remaining un-routed stop, select one un-routed stop and select an associated position on a current route that yields the best insertion cost. If the new route duration satisfies the maximum workday duration limit, then execute the insertion and repeat this step; otherwise, go to step (1);
  • [0335] (3) terminate the heuristic.
    Based upon this insertion framework, the construction of a new route starts by selecting a seed stop.

[0336] In one embodiment of the present invention, the strategy used is to select the un-routed stop located furthest from the depot or hub 30. The insertion cost may be calculated as follows. Assume the current route consists of stops (1, 2, . . . , i, j, . . . , n), in that order. Stop u is an un-routed stop. The cost of inserting stop u between stops and j may be expressed as:
c(i,u,j)=d_iu+d_uj−d_ij
where d is the distance matrix among stops. If this insertion is not feasible, then the insertion cost c is infinite.

[0337] The best insertion position for stop u, (i*,j*) is the one that yields:
c(i*,u,j*)=min_{k=1, . . . , n}{c(k,u,k+1)}

[0338] The next stop to be inserted into the route, then, is the one that yields:

[0339] $c\left(i^{**},u^{*},j^{**}\right)=\underset{u}{\min }\left\{c\left(i^{*},u,j^{*}\right)\right\}$

[0340] and stop u* is inserted between stops i** and j**.

[0341] A parallel route construction heuristic, for example, may begin with k routes, each having a seed stop, such that the initial route may be expressed in the form of (depot, seed, depot). The value of k may be determined from the result of the sequential route construction heuristic, above. The steps in the parallel route construction heuristic may be similar to those in the sequential route construction heuristic. In each step of the parallel route construction heuristic, select an un-routed stop to be inserted into a position on the route that yields the best insertion cost. If there are no un-routed stops remaining, the algorithm terminates.

[0342] In one embodiment, the parallel route construction heuristic may employ a regret measure to select the next un-routed stop to be routed. The regret measure may be based on the insertion cost function c(i,u,j) mentioned above in the sequential heuristic. For each un-routed stop u, the algorithm first calculates the feasible insertion position of u on each of the k routes. The best insertion cost for stop u may be expressed as c_r(u)=c(i_r*,u,j_r*), for a best insertion position located between stops i_r* and j_r*. Applying these conditions, the best insertion cost overall for stop u is the one that satisfies the equation:

[0343] $c^{*}\left(u\right)=\underset{r=1,...,k}{\min }c\left(i_r^{*},u,j_r^{*}\right)$
If route r* is denoted as the route on which c_r(u)=c*(u), then the regret cost for stop u may be expressed as:

[0344] $\mathrm{regret}\left(u\right)=\sum _{r\ne r^{*}}\left(c_r\left(u\right)-c_{r^{*}}\left(u\right)\right)$

[0345] The algorithm then selects the next un-routed stop u* that has the maximum regret cost:

[0346] $\mathrm{regret}\left(u^{*}\right)=\underset{u}{\max }\left(\mathrm{regret}\left(u\right)\right)$
and inserts stop u* into its associated best route r*, at the best position (i_r**, j_r**).

[0347] The algorithm repeats the above process until all stops are routed or there is no route that can accept an un-routed stop, at which point a new route is created following the same procedure as described in the sequential route construction heuristic.

[0348] In one embodiment, the system of the present invention provides three types of route improvement procedures:

• • [0349] (1) The Intra-route Exchange procedure attempts to re-order the stop sequence within a given route in order to achieve better route cost. In one embodiment, the best route cost has the shortest total route duration, within the maximum workday duration constraint. One of the methods used to accomplish the Intra-route Exchange procedure may be an algorithm known as a two-opt exchange heuristic, which involves replacing two arcs of a route with two other arcs that are not on the route.
  • [0350] (2) The Inter-route Transfer procedure removes a stop from one route and inserts it into another route, to determine if the transfer reduces the total route cost.
  • [0351] (3) The Inter-route Exchange procedure swaps a stop on one route with a stop on another route, to determine if the exchange reduces the total route cost, within the maximum workday duration constraint.

[0352] 9.2. Adaptation of a VRP Algorithm

[0353] The operational cell routing model of the present invention is expressed in a three-index formulation. The inventive model is similar to a classical vehicle routing problem (VRP) constrained by a maximum driver workday duration, except for the pre-assignment constraint in Equation (20) above and the following three differences:

• • [0354] (1) the inventive model is for routing cells instead of individual stops;
  • [0355] (2) the inventive model incorporates the fact that some of the cells are assigned as core cells 60; and
  • [0356] (3) the inventive model includes the effect of driver learning.
    In one embodiment, the method of the present invention addresses these three differences, as discussed below.

[0357] First, in one embodiment, the present invention includes a VRP solution method adapted for routing entire cells instead of individual stops. In one aspect, a cell may be treated mathematically as a large super-stop, having a location at the centroid of all stops within the cell and a service time equal to the time needed to serve all the stops within the cell including travel time between stops.

[0358] For a route consisting of cells (1, 2, . . . , i, j, . . . , n), in that order. Cell u is an un-routed super-stop. The cost of inserting cell u between stops i and j may be expressed as:
C(i,u,j)=D_iu+D_uj−D_ij
where D is the distance matrix among cells. If the insertion is not feasible, the insertion cost C is infinite. The cost C may be referred to as a cost constraint, in relation to the adapted VRP algorithm.

[0359] Second, in one embodiment, the present invention includes a VRP solution method adapted for the fact that some of the cells are assigned as core cells 60 to a regular driver. The number of core cells k may be determined by the number of regular drivers or by the expected minimum number of drivers needed each day. Therefore, k partial routes may be used to serve as a starting point. The series of k partial routes may be referred to as a core constraint, in relation to the adapted VRP algorithm being formulated. The starting point of k partial routes is similar to the starting point for the parallel route construction heuristic described above. Accordingly, there is no need to run a sequential route construction heuristic to determine the number of routes.

[0360] Because the core cells 60 only form an un-sequenced partial route, the cells 60 must be re-ordered with the objective of minimizing the total route duration. In application, this is similar to a traveling salesman problem (TSP), where the unit is a cell instead of a single stop. In one embodiment, the method of the present invention includes a simple insertion heuristic for solving this problem. The insertion cost may be calculated as shown above.

[0361] In another aspect, the method of the present invention makes certain the core cells 60 are not moved or transferred to other routes and drivers by the route improvement procedures, such as inter-route transfer and inter-route exchange.

[0362] Third, in one embodiment, the present invention includes a VRP solution method adapted to include the effect of driver learning. The route construction method may include the step of building a dynamic driver performance matrix Pmat, which may be used to store each driver's historical performance level in each cell, in terms of the percentage of standard time the driver needs to serve each cells and travel from cell to cell. In the matrix, Pmat (i, k) may represent the current performance level for a driver k in a cell i. After a route plan is generated for a current day, the matrix Pmat may be updated with the day's cell assignments.

[0363] If cell i is assigned to driver k on a subsequent day, then according to learning curve, Pmat (i, k) should be updated to:

[0364] $\mathrm{Pmat}\left(i,k\right)=\max \left\{T_0·{\left({\left(\frac{\mathrm{Pmat}\left(i,k\right)}{T_0}\right)}^{-\frac{1}{f}}+1\right)}^{-f},T_{\infty }\right\}$
where T₀ is the starting performance level, f is the learning and forgetting rate, and T_∞ is the learning limit. Beginning from T₀=100%, then:

[0365] $\mathrm{Pmat}\left(i,k\right)=\max \left\{{\left({\left(\mathrm{Pmat}\left(i,k\right)\right)}^{-\frac{1}{f}}+1\right)}^{-f},T_{\infty }\right\}$
If cell i is assigned to a driver other than driver k, then according to the forgetting curve, Pmat (i, k) should be updated to:

[0366] $\mathrm{Pmat}\left(i,k\right)=\min \left\{T_{\infty }·{\left({\left(\frac{x}{T_{\infty }}\right)}^{\frac{1}{f}}+1\right)}^f,T_0\right\}$
This dynamic driver performance matrix Pmat serves as a multiplier for the total time T_ik needed to serve all stops in cell i by driver k, and the travel time D_ij from cell i to cell j. Applying this multiplier means that, wherever the terms T_ik, and D_ij appear, the new terms Pmat(i,k)*T_ik and Pmat(i,j)*D_ij are used instead. The new terms reflect the time savings generated by driver learning. Thus, the mathematics confirm the advantage of assigning each cell to the driver who is most familiar with it. The new Pmat terms may be referred to as a driver learning constraint, in relation to the adapted VRP algorithm being formulated. In one aspect, the method of the present invention used to build the dynamic driver performance matrix Pmat is a heuristic method because it is myopic, in the sense that it doesn't consider the effect of the driver assignment on learning over the long term.