Dynamic vehicle routing using hybrid genetic algorithms

(1)

Proceedings of the 1999 EEE International Conference on Robotics & Automation

Detroit, Michigan May 1999

Dynamic Vehicle Routing Using Hybrid Genetic Algorithms

Wan-rong Jih

jih@robot .csie.ntu.edu.tw

Jane Yung-jen Hsu

yjhsu@csie.ntu.edu. tw

Department of Computer Science and Information Engineering

National Taiwan University

Taipei, Taiwan

Abstract

This paper presents a novel approach to solving the single-vehicle pickup and delivery problem with time windows and capacity constraints (or single-vehicle P D P T W ) . While dynamic programming has been used to find the optimal routing to a given problem, it re- quires time exponential in the number of tasks. There- fore, at often fails to find the solutions under real-time conditions in an automated factory. This research explores anytime problem solving using genetic algo- rithms. By utilizing optimal but possibly partial solu- tions f r o m dynamic programming, the hybrid genetic algorithms can produce near-optimal solutions for problems of sizes up to 25 percent bigger than what can be solved previously. This paper reports the exper- imental results of the proposed hybrid approach with f o u r different crossover operators as well as three mu- tation operators. The experiments demonstrated the advantages of the hybrid approach with respect to dy- namic task requests.

1 Introduction

To this day, only relatively small VRP instances can be solved to optimality.

In this paper, we consider the single-vehicle pickup and delivery problem with time windows and capacity constraints (single-vehicle PDPTW), which will be formally defined in section 2. Most exact algorithms capable of finding optimal solutions to this problem are based on dynamic programming. For a single- vehicle PDPTW with n tasks, the computational com- plexity of dynamic programming has been shown to be bounded by (an

+

1)23"[5]. Therefore, existing algorithms cannot satisfy real-time requests.

In order to handle dynamic task requests efficiently, it is desirable to have heuristic algorithms that can generate (sub-optimal) solutions to large problems on

demand. The solutions should be improved incrementally whenever more time becomes available. Genetic algorithms (GAS) can generate incrementally better solutions at each succeeding generation regardless of the problem size. Therefore, they provide a reasonable solution to the dynamic single-vehicle PDPTW problem. Unfortunately, the solutions from simple GAS are often of poor quality.

This research explores hybridizing genetic algorithms with dynamic programming in order to take advantage of the complementary properties of both in solving the dynamic single-vehicle PDPTW problem. The proposed hybrid approach starts by using dynamic programming to generate the optimal routes. If optimal solutions are not found within the specified time slot, the partially constructed routes are passed to the genetic algorithms. The sub-routes provide the basis for generating an initial population that often leads to better convergence than a randomly generated initial population. The hybrid approach enables dynamic programming to achieve real-time performance, and improves genetic algorithms in approx- imating near-optimal solutions.

The next section introduces the single-vehicle PDPTW, and section 3 presents the hybrid approach. The experimental results and analysis are shown in section 4, followed by the conclusion in section 5.

2 The Single-vehicle PDPTW

We should know the static single-vehicle PDPTW before we focus our attention on the dynamic single- vehicle PDPTW. The section 2.1 will describe the static case of single-vehicle PDPTW. The dynamic case will be presented in section 2.2.

2.1 Static case

Suppose that N = { 1,.

. .

,

n } is a set of tasks, where

n is the number of tasks. To accomplish each task

(2)

i E N, the vehicle should pick up the goods at the pickup location if and then transport the goods to the delivery location i-. According to these pickup and delivery locations, V + = {i+

I

i E N} and

V - = {i-

I

i E N} indicate the set of pickup locations and the set of delivery locations respectively. Let G = (V, A) be a graph where V = (0) U V + U V-

is a set of vertices and A = { (r, s)

I

r

#

s, r, s E V}

is a set of arcs. Vertex 0 represents an initial depot of the vehicle. Each arc ( r , s ) E A is associated with a non-negative travel time d,,, and each task i E N is associated with a non-negative demand qi. The

total demand of a vehicle route may not exceed the vehicle capacity Q. Specifically, for every task i are required for picking up the goods within the time interval [ai+, bi+], and transporting the goods to its de-

livery location i- on the time interval [ u i - , b*--]. In addition, let t , and 1, represent the arrival time and the load of a vehicle respectively, when the vehicle arrives at a location r E V.

Our goal is to find a vehicle route starting from an initial depot, finishing all the requests of tasks, and ending at one of the delivery locations. According t o the route, both the total traveling time and waiting time of the vehicle are minimized. Certainly, this route must be a feasible route that is satisfied the capacity

and the time windows constraints of each task. In addition, the feasible routes visit the pickup location

i+ before the delivery location i- of task i E N .

Initially, we assume that the vehicle is stopped at an initial depot 0. Either the pickup location or the

delivery location are associated with an advanced time

window [a,, b,], where r E V + U V - . The weight of

goods q, for a task i E N is known, when the task i

is assigned. We assume the value of travel time d,, is

already known, where r , s E V. If the vehicle arrives

at a location r E V + U V - and its arrival time t , is earlier than a,, it has t o wait. Every task i E N is subject to a,+

5

bi+, ai-

5

bi- and bi- - ai+ 2 di+,-.

The solution cost of the static single-vehicle PDPTW problem is minimized.

Z o b j e c t i u e = w1 drsxrs

+

w2 f w a i t i n g ( r )

T,SEV r e V + U V -

(1)

The notation x,,, r, s E V, is defined as

{

0 , otherwise.

1, if arc (r, s) is used by vehicle;

2,s =

The function f w a i t i n g ( r ) , defines the waiting time of a

location r E V + U V - as

a,

-

t,, if vehicle arrives at location r early;

f w a i t i n g ( r ) =

The first operand of the equation (1) is the total traveling time needed by the vehicle for completing the route, and the second operand of equation (1) is the total waiting time. The two factors 01 and wp are the weights, which reflect the relative importance of these two parts. We generally set w 1

+

w 2

>

0 and W l , W 2

2

0.

Solutions to the single-vehicle PDPTW are subject to the following constraints[2]. First, the vehicle starts from an initial depot, and each location is visited only once. Second, each task r E N should be picked up before it is delivered, which is called precedence con- straints. Third, the total load allocated to a vehicle cannot exceed its capacity. This constraint is named

capacity constraints. Fourth, whenever the vehicle ar-

rives at location i E V + U V - at time ti, the criterion

ti

5

b, must be satisfied. In general, we named the

fourth constraint as time window constraints. Finally,

the vehicle routes will be open paths, ending at any one of the delivery locations.

2.2 Dynamic case

In the static cases, tasks to the problem do not change, either during the execution of the algorithm, or during the eventual execution of the route. By con- trast, in dynamic cases, tasks may change during the execution of the algorithm and the eventual execution of the route. For the requirements of a dynamic vehicle routing problem, we should modify the objective function Z o b j e c t i v e to accomplish a route in real time.

A modified objective function is defined as

Z d y n a m i c = Z o b j e c t i v e

+

[a1 f d e l a y ( r ) r E V + U V -

+ a 2 f o u e r l o a d ( r ) ] r E V + u V -

The objective function Z d y n a m i c is based on the objec-

tive function of the static case, but with some penal- ties. For any r E V + U V - , the function f d e l o y ( r ) and

foverload(T) represent the vehicle delay time and over-

load at location r respectively, where a 1 and a2 are penalty coefficients.

The functions fdelag(r) and foverload(r) are defined

as following (r E V + U V - ) :

t , - b,, if a vehicle arrives at location r lately;

otherwise.

if the current load 1, ex- ceeds the vehicle capacity

Q ;

{

0, f d e t a y ( r ) =

I

0, otherwise. 454 otherwise.

(3)

According to the definition of Z d y n a m i c , we will try

to minimize it.

3 Hybrid Approach

Figure 1 shows the architecture of the hybrid genetic algorithms.

module

Re-Planned DP

module Illod"k

Figure 1: Architecture of hybrid GA

3.1 Pre-planned module

Algorithm 1 Genetic algorithm of single-vehicle PDPTW

1: t = 0;

2: Generate a population P ( t ) ;

3: Use Z d y n a m i c to evaluate chromosomes in P ( t ) ; 4: while termination condition not satisfied do 5:

6: 7:

8: Perform crossover; 9:

Initialize a temporary population PI; for i = 1 to IP(t)I do

Select two parents from P ( t ) ;

if the offspring and its parents are identical then 10: Perform mutation; 11: end if 12: 13: end for 15: Replace P ( t ) by PI; 16: end while

17: Output the solutions;

Place the offspring into PI; 14: t = t -I- 1;

A pre-planned module will arrange the tasks and prepare the information for the DP module. The pre- planned module collects the tasks that have not been finished, and set a new initial depot of the vehicle. The execution time and the load of vehicle are not equal to zero and the pre-planned module should maintain the proper values of them.

3.2

DP

module

The DP module performs a dynamic programming algorithm. When a specific time is expired, the DP module will pass its unfinished sub-routes to a temporary result pool; these unfinished sub-routes will be the initial population of genetic algorithms. We adopt the dynamic programming approach that proposed by Psaraftis[4]. In the current state of knowl- edge, the single-vehicle dial-a-ride problems can rarely be achieved to optimization when the number of tasks is more than 40. As to the dial-a-ride problems, they might be a special case of PDPTW when every task demands of the PDPTW are equal. Detailed algorithms of dynamic programming can be found in many publications[4, 5, 31.

3.3 Genetic algorithm

We summarize the genetic algorithm used in our work as Algorithm 1. The procedure of step2 will generate an initial population. The initial population of the hybrid genetic algorithms is based on the partially constructed routes of DP module; the sub-routes are in front of the initial routes. According to the step 7, we apply the tournament selection to select two parents,

and the tournament size is 2. Four crossover operators

and three mutation operators will be discussed in this paper. The algorithm will be executed until one of the following conditions is met: (1) The number of generation is larger than the maximum generation. (2) The solution that has the best fitness value appears more than a specific sequence. (3)

<

0.005. The a ( P )

and m ( P ) are the standard deviation and the mean

of the fitness value in the population P respectively.

Currently, we set the values of maximum generation and the appearence sequence are equal to 2000 and 100 respectively.

m ( P ) -

3.3.1 Representation

Given a single-vehicle PDPTW with n tasks, a so-

lution is encoded as a chromosome, which is rep- resented by a permutation of locations from 1+ to

n-. For example, in 2 tasks problem, if a route is 0

+

1+

+

2+

+

2- t 1-, the representation of the chromosome will be (1+ 2+ 2- 1-). This representation could be understood easily and the space require- ment is less than other representations. In addition, it is convenient for implementation.

In the single-vehicle PDPTW, the precedence constraints can not be violated. The representation does not preclude routes from violating the precedence constraints. A simple algorithm is used to maintain the feasibility of the corresponding routes. A chromosome adjustment algorithm will exchange the positions of

i+ and i-, when the vehicle visits the delivery location i- before the pickup location i+ of a task i E N .

(4)

somes will be satisfied the precedence constraints by making the pickup location be appeared before the delivery location for any specific task.

3.3.2 Crossover operators

In the present work, we compare to the performance of four crossover operators on the single-vehicle PDPTW problem. These operators are order crossover (OX), uniform order-based crossover (UOX), merge cross #1

(MX1) and merge cross #2 (MX2)[1]. The first two operators are traditional crossover operators and the last two operators use a global precedence vector to be the guidance of crossover. We will not describe the two traditional crossover operators. The other two crossover operators are described below.

Most traditional crossover operators for order- based GAS do not have strong connection t o the constraints of the problems they are applied. If we apply the traditional crossover operators to the single-vehicle PDPTW, they might not be conducive t o the search- ing process of optimal solutions because the information of the constraints is not used by these operators. On the other hand, the notion of merge crossover operators is that a global precedence among genes inde- pendent of any chromosome, rather than the behav- ior of traditional crossover operators that defines a local precedence among genes specific t o a chromosome. That is, each gene in the chromosome has a precedence relationship to every other gene. From the charac- teristics of constraints of the single-vehicle PDPTW, a global precedence relation probably exists among genes. The global precedence vector is formed by such relationship and it could be the offspring-generating guidance.

In our single-vehicle PDPTW, each gene is either a delivery point or a pickup point and has an associated time window. Therefore, we can make use of the time window [a,, bp] of the location r E V + U V - and a precedence relationship among each earliest processing time a,. Given r,s E V + U V - , if a, less than a,, let the location r be appeared before the location s

in the global precedence vector because it will be a

reasonable solution to serve the location r before the location s. In other words, we can sort all a, in the ascending order, and the sorting result will be a global precedence vector.

We demonstrate the operations of MX1 with an example. Suppose that the set of tasks

problems consisting of 10, 20, 30, 40 and 50 tasks, respectively. In these problems, the pickup locations and the delivery locations are randomly placed in a rectangular grid. The locations of each task are associated with a time window, which is generated randomly. The Euclidean distance is used to measure the distance between a pair of locations.

In our experiments, the global precedence vector of

MX1 and MX2 is based on the lower bound of time windows. We test the four crossover operators associated with three different crossover rates, the rates are

0.45, 0.6 and 0.75. The population size of each problem is the number of task multiplied by 5. That is, if the number of task is 40, the total number of location is 81 and the population size is 200. The value of each factor in the objective function will be a1 = 10, a2 = 50, w1 = 1, and w2 = 2.

At

first, we would like to make a comparison among the crossover operators without the mutation operators, in order t o study the pure relative performance of the crossover operators. In Figure 2, the mutation operators are not applied to the single-vehicle PDPTW problem. The solution cost is based on Zobjective

where w1 = w2 = 1. In Figure 2, for each generation, the values of solution cost are the average in 20 runs.

Figure 2 shows the MX1 could achieve better solution cost than the other crossover operators do. Ac- cording to the Figure 2, OX is not a proper operator for the single-vehicle PDPTW problem. OX creates offspring which inherits sub-tours of its parents. Such a new route is analogous to its parent routes. The solutions were trapped to similar routes, and therefore

0.45 1.000

1.000

~ ‘ 0 0 0

j

1.~000 0.60 1.000 1.000 1.000

1

1.000 1135

4

20 . . . ... 0.45 1.000 1.000 1.000 1.000 0.60 1.000 1.000 1.000 1.000 1794 0.75 1.000 1.000 1.000 1.000 0.45 1.033 1.000 1.000 1.033 I 0 50 150 m 250 300 Irn generation

Figure 2: The comparison of crossover operators

0.75 0.45

feasible solutions may not be found. As to MX2, a location with the lowest precedence is moved to near the end of the routes, and the solution may be stuck to a local minimum. The solution cost of UOX is approximate to the solution cost of MX1 when the generation is larger than 200, but MX1 achieves the same solution cost before the 50th generation. For the real-time requests, the UOX may not be suitable for the dynamic single-vehicle PDPTW problem.

I I 2602

1

40y2

’

0.75 1.040 1.000 1.000 1.051 0.45 5336 5247 5247 5336 0.75 5341 5266 5276 5336

(6)

30 and 40 tasks are listed in the last column of Ta-

ble l. The mutation operator l-’ means that no mutation operator is applied to the genetic algorithms. As to the problem of 50 tasks, the results are shown

in the last three rows of Table 1. Most of exact algorithms applied to vehicle routing problems can not solve relatively larger problems, and we can not obtain the optimal values neither. From the results of Table 1, it shows the mutation1 and mutation2 attain better solution cost than the other mutation operator does.

Figure 3 shows the average solution costs of the

hybrid genetic algorithms are better than the average costs of the traditional genetic algorithms. A dynamic

Figure 3: DP+GA vs. GA only

programming is applied to generate the initial population instead of creating the it randomly. Since the initial population filtered out some of the infeasible routes and put the initial population contents more feasible sub-routes, the hybrid genetic algorithm al- ways obtain the lower solution cost than the traditional genetic algorithms.

5 Conclusion

In this paper, we show that the hybrid genetic algorithm solves the dynamic single-vehicle PDI’TW problem. With such approach, that takes advantage of both dynamic programming and genetic algorithms. The approach enables dynamic programming to achieve real-time performance and genetic algorithms to approximate optimal solutions. The initial population created by the dynamic programming instead of generating it randomly. The dynamic programming passes the unfinished routes to genetic algorithms to accomplish the real-time performance. A good initial population of the genetic algorithms is improved by the dynamic programming.

The hybrid approach can generate incrementally better solutions at any time, which is essential for deal- ing with dynamic task requests. Comprehensive experiments using an efficient and flexible implementation of the hybrid genetic algorithms were performed. The experimental results showed that the hybrid approach could produce near-optimal solutions for problems of sizes up to 25 percent bigger than what can

be solved previously by dynamic programming. In addition, the hybrid approach may find sub-optimal solutions for dynamic vehicle routing problems of any size.

References

[l] J. L. Blanton Jr. and R. L. Wainwright. Multi- ple vehicle routing with time and capacity con-. straints using genetic algorithms. In Proceedings of the Fifth International Conference on Genetic Al- gorithms and Their Applications, pages 452-459, 1993.

[2] M. Desrochers, J. K. Lenstra, and M. W. P. Soumis. Vehicle routing with time windows: Opti- mization and approximation. In B. L. Golden and

A. A. Assad, editors, Vehicle Routing:Methods and Studies, pages 65-84. Elsevier Science Publishers,

North-Holland, Amsterdan, 1988.

[3] J. Desrosiers, Y. Dumas, and F. Soumis. A

dynamic programming solution of the large-scale single-vehicle dial-a-ride problem with time windows. American Journal of Mathematical and Management Sciences, 6(3 & 4):301-325, 1986. [4] H. N. Psaraftis. A dynamic programming solution

to the single vehicle many-to-many immediate re- quest dial-a-ride problem. Transportation Science, 14(2): 130-154, 1980.

[5] H. N. Psaraftis. An exact algorithm for the single vehicle many-to-many dial-a-ride problem with time windows. Transportation Science, 17(3):351-

357, 1983.