## Making the Yellow Angels Fly: Online Dispatching Of Service Vehicles in Real Time

**May 3, 2002**

Figure 1. Partial snapshot of a real-world instance of the ADAC-problem (gray: units; black: events). Lines indicate currently planned tours.

**Martin Grötschel, Sven O. Krumke, Jörg Rambau, and Luis M. Torres**

Combinatorial online optimization is an area that has lots of applications and abundant potential for significant progress, both in theory and in practice. In this short note we sketch a typical large-scale online optimization problem: dispatching of service vehicles. We discuss some theoretical and practical issues that arise, and explain, very briefly, how we approach this problem mathematically. Online problems are a battlefield of heuristics, with many strong claims advanced about the quality of the solutions obtained. As shown below, careful modeling and the use of more mathematics can produce superior results.

[*On Wednesday, July 10, as one of five invited plenary speakers at the SIAM 50th Anniversary and 2002 Annual Meeting (Philadelphia, July 8-12), Martin Grötschel will discuss the "tremendous progress" of the last fifty years in the significantly overlapping fields of discrete mathematics and theoretical computer science, with a focus on large-scale real-world examples of the type presented here.-ed*.]

**The Problem**The German automobile association ADAC (Allgemeiner Deutscher Automobil-Club), second in size only to the AAA, the largest such organization in the world, maintains a heterogeneous fleet of more than 1,600 service vehicles to help drivers whose cars break down on the road. All ADAC service vehicles have a distinctive yellow color, and the help provided by the experienced mechanics often appears to have been sent from heaven; for these reasons, the service vehicles are affectionally called "yellow angels." In this article we are a little more down-to-earth and just refer to them as "units."

ADAC runs five help centers (Pannenhilfezentralen), which are located throughout Germany and operate 24 hours a day, seven days a week. (Human) dispatchers have to reply to help requests ("events," for short) instantly. Their task is to assign a unit (equipped with GPS) to serve each customer and to predict the estimated time of arrival at the customer's location.

In addition to the ADAC fleet, about 5,000 units operated by service contractors can be employed to cover events that could not otherwise be served in time. In the continuously running planning process, a dispatcher can revoke a particular assignment to an ADAC unit, provided that the unit has not yet reached the customer.

There is no unique objective. The goals are high-quality service (e.g., short waiting times for the customers, preferably well below an hour but depending on the system load) and low operational costs (through measures like short total tour lengths and limited overtime expenditures). Figure 1 (page 1) shows a typical, obviously not uniform distribution of units (gray) and events (black) in a real-world instance of ADAC.

With increasing costs and request volume, ADAC's manual dispatching system has come under stress. The challenge is to design an automatic online dispatching system that will guarantee small waiting times for events and that will keep operational costs low. The modeling of such a task is not easy, as many technical and organizational side constraints---some hard, others soft---have to be considered, and it takes some time to figure out which constraints and objectives really count. We do not describe the complete model here but consider, as one example of a management decision, ADAC's imposition of a soft deadline on the service time of an event: The deadline can be missed at the cost of a linearly increasing lateness penalty (soft time windows).

In the remainder of this article we refer to the dispatching problem sketched above as the "ADAC-problem."

**The Problem with the Problem**In fact, we face two problems. The first is the task of computing an optimal dispatch for a "snapshot" situation (that is, a dispatch that is optimal for a given set of events and contractors at some moment in time). Because it contains the classic vehicle-routing problem as a special case, this problem is NP-hard. The difficulty is that in the dynamic planning process, a dispatch now has to be returned in no more than 15 seconds (the "real-time problem").

A problem of this type can be described as a multidepot vehicle-routing problem with soft time windows. To date, most research in this area has considered hard time windows (see [1] and references therein for a survey of various problem types and exact algorithms; see [2] for a discussion of recent improvements in the efficiency of exact algorithms). For instant answers, the literature seems to suggest that use of meta-heuristics is unavoidable ([4] describes a tabu search approach for problems with soft time windows; [3] presents an approach based on genetic algorithms). These meta-heuristics, however, do not provide evaluations of their results, a significant drawback in our opinion.

Integer programming, the approach we prefer (maybe in conjunction with whatever heuristics turn out to be useful), often yields lower bounds on the costs that are unavoidable with any solution method. It was this lower-bound argument that convinced the ADAC not to simply pay one of the many commercial providers of meta-heuristic software, but to let us look a little deeper into the problem.

The second problem we face is more subtle: We do not know where future events will pop up (the "online problem"). A decision that is "optimal" at some point in time can prove later to have been unwise. In particular, even if we were able to compute locally optimal dispatches for any snapshot situation, we would not necessarily obtain a dispatch that is (in hindsight) optimal for the whole planning period.

**Toward a Solution**A closer inspection of the problem and typical instances leads us to expect that "snapshot optimization" does have potential. Time windows are helpful. Although they often seem to make routing problems harder, the contrary is true in our online situation. We compute optimal snapshot-dispatches whenever new events occur and use those results for subsequent routing decisions until the arrival of new events. In planning tours for the service units, we have each unit end its tour at its home base; in this way we avoid units' drifting far from home and requiring expensive return trips at the end of their shifts.

For the real-time problem, the ADAC-problem luckily turns out to have some structure that we can use to our advantage. Because it takes some time for a service unit to deal with an event and because the time windows for events are quite tight, tours in an optimal solution cannot be too long (at most four to five events per unit). So, again because of the time windows, chances are good that all relevant tours can be enumerated implicitly.

What comes to mind at this point is a well-known column-generation and set-partitioning approach: Generate all relevant tours and partition the set of events into a subset of tours, with some single events outsourced to the contractors. The fractional set-partitioning solution (LP relaxation) serves as a lower bound (unavoidable costs) for a snapshot-dispatch. This is basically what is done in most of the known exact algorithms for vehicle routing. Our task is to turn the known methods-which usually require substantial time for warmup or even preprocessing---into an algorithm that delivers provably good snapshot---dispatches within seconds.

We accomplish this with a dynamic pricing control in the column-generation procedure. Early in the optimization, we work on an extremely restricted search space for new tours, using a mix of heuristics to find new tours of restricted depth. The search space for new tours is extended as the estimates for the dual prices in the LP relaxation of the set-partitioning problem get closer to the truth.

**Results**Our implementation of the column-generation method ZIBDIP (ZIB Dispatching via Integer Programming) was competing with an experimental meta-heuristic prototype programmed for this particular purpose by ADAC's software provider, Intergraph Public Safety (IPS). The evaluation was done on varying amounts of snapshot data. The results were conclusive: In this special application, the integer programming approach completely outperforms the meta-heuristic paradigm. ZIBDIP produced optimal (or optimal to within 1%) solutions for all real-world examples (up to 215 events for 98 units) within one to twenty seconds on a standard 800-MHz PC; the meta-heuristic, by contrast, was unable in several instances to get closer to optimality than 25%. In larger, artificially generated problems (up to 775 events for 215 units), this trend was vastly amplified.

The superb performance on real data convinced ADAC and IPS that our algorithm is the way to go for the snapshot-dispatches, and IPS agreed to implement the production software based on the ZIBDIP algorithm.

With the real-time issue solved to everyone's satisfaction, we turned our attention to the online issue. What quality would the repeated generation of optimal snapshot-dispatches deliver over time?

Although development of the final version of the online algorithm is still under way, we have run online tests comparing two strategies: planning snapshot-dispatches after each new job or at fixed frequencies (one to five minutes). The trend was slightly in favor of planning after each new job.

The results were promising: On average, the online costs were about 40% above the lower bound---not too bad, in our experience, for an online algorithm. We hope, however, to improve on these figures by utilizing estimates of future events in the snapshot-dispatches. This is work in progress and leads to most interesting questions as to how knowledge about the future can be exploited in combinatorial online optimization.

**References**[1] J. Desrosiers, Y. Dumas, M.M. Solomon, and F. Soumis,

*Time constraint routing and scheduling*, in

*Network Routing, Handbooks in Operations Research and Management Science*, M. Ball, T. Magnanti, C. Monma, and G. Newhauser, eds., 8, Elsevier, Amsterdam, 1995, 35-140.

[2] J. Larsen,

*Vehicle routing with time windows-finding optimal solutions efficiently*, DORSnyt (newsletter of the Danish OR society; in English), 116 (1999).

[3] S.J. Louis, X. Yin, and Z.Y. Yuan,

*Multiple vehicle routing with time windows using genetic algorithms*, 1999 Congress on Evolutionary Computation, Piscataway, NJ, IEEE Service Center, 1999, 1804-1808.

[4] E. Taillard, P. Badeau, M. Gendreau, F. Guertin, and J.-Y. Potvin,

*A tabu search heuristic for the vehicle routing problem with soft time windows*, Transportation Sci., 31 (1997), 170-186.

*Martin Grötschel is a professor at TU Berlin and vice president of the Konrad-Zuse-Zentrum für Informationstechnik Berlin (ZIB); Sven O. Krumke and Jörg Rambau are researchers at ZIB, and Luis M. Torres is a PhD student.*