Вы находитесь на странице: 1из 40

TRB Paper No.

03-4158

A Stepwise Highway Alignment Optimization Using Genetic Algorithms by


Eungcheol Kim, Ph.D. Research Fellow Department of Highway Research The Korea Transport Institute (KOTI), South Korea TEL: +82-31-910-3057, FAX: +82-31-910-3235 E-MAIL: eckim@koti.re.kr

Manoj K. Jha, Ph.D., P.E. (Corresponding Author) Assistant Professor Department of Civil Engineering Morgan State University 5200 Perring Parkway Baltimore, MD 21251 TEL: 1-443-885-1446, FAX: 1-443-885-8218 E-MAIL: mkjha@eng.morgan.edu and Bongsoo Son, Ph.D. Associate Professor Department of Urban Planning and Engineering Yonsei University, South Korea TEL: +82-2-2123-5891, FAX: +82-2-393-6298 E-MAIL: sbs@yonsei.ac.kr

November 2002

Submitted for presentation at the 2003 Annual Meeting of the Transportation Research Board and for publication in the Transportation Research Record

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

ABSTRACT
In this paper we propose a stepwise highway alignment optimization approach using genetic algorithms for improving computational efficiency and quality of solutions. Our previous work in highway alignment optimization has demonstrated that computational burden is a significant issue when working with a Geographic Information System (GIS) database requiring numerous spatial analyses. Furthermore, saving computation time can enhance adoptability of a model especially when a study area is relatively large or involves many sensitive properties or if locating complex structures such as intersections, bridges and tunnels is necessary. It is well acknowledged that in many optimization processes subdividing large problems into smaller pieces can decrease the computation time and produce a better solution. In this research two different population sizes are used to develop a stepwise alignment optimization when employing genetic algorithms in suitably subdivided study areas. An example study shows that the proposed stepwise optimization gives more efficient results than the existing methods and also improves quality of solutions.

Key Words: Stepwise optimization, Genetic algorithms, Computational efficiency, Highway


alignment optimization, Geographic information systems, Segmentation

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

INTRODUCTION
The highway alignment optimization involves finding the best highway alternative between a pair of points (1-5). The problem can be stated as follows: Given two end points in the study area and allowing the existing conditions of the study area changeable, find the best alignment among alternatives to optimize a specified objective function, while considering needed structures and satisfying design and operational requirements. For more reliable and realistic applications highway alignment optimization processes should consider many factors, which increase the complexity of the problem. The factors may include structures, topography, socio-economics, ecology, geology, soil types, land use patterns, environment and even community concerns. They are considered with different emphasis and levels of detail at different stages in the alignment selection processes. Traditionally, these processes have consumed much time and effort of agencies, planners, engineers and residents. Several models have been developed in response to this need. They can save considerable time and costs compared to the traditional manual methods using computers and mathematical formulations (6-8). Recently, a solution approach (1, 4-5) based on genetic algorithms (GAs) for three-dimensional highway alignment optimization has been developed. The GA advantages to the highway alignment optimization problem over traditional methods have been extensively covered in (1-5); therefore, have been skipped here for brevity. A model integrating geographic information systems (GIS) with such a GA has also been developed (2). Furthermore, there has been an effort to incorporate structures such as intersections, tunnels, bridges and interchanges into the optimization process for improving practical usability of the models (3).

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

Although the first objective of the many developed models is to obtain the best alignment (global optimum or at least near global optimum), computational efficiency of the models is also of great concern since it largely affects the degree of a models adoptability. The computational burden especially increases (2) when the number of properties to be analyzed for right-of-way cost calculation and environmental impact assessment increases. It is well known that in many optimization processes, subdividing large problems into suitable pieces can decrease the computation time and produce a better solution. This argument also applies to this study, since optimizing highway alignments repeatedly involves fine-tuning search steps during successive search processes.

LITERATURE REVIEW
Theoretically, highway alignment optimization problem involves an infinite number of alternatives to be evaluated. In previous applications (1-5) the optimization problem was formulated as a cost minimization problem in which cost functions are non-differentiable, noisy and implicit. Thus, it is inevitable to use fast and efficient search algorithms to solve such a problem. According to Table 1, seven search methods (1-28) are used for alignment optimization models. Among those, all have some critical defects when applied to the highway alignment optimization problem except genetic algorithms. Table 2 summarizes these defects.

GENETIC ALGORITHMS AS AN OPTIMAL SEARCH Genetic Algorithms (GAs) have been proven to be very effective to highway alignment optimization problems (1-5) since they can effectively search in a continuous search space without getting trapped in local optima. Goldberg (29) states four important distinctions of GAs

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. over other search methods: (1) GAs work with a coding of the parameter set, not the parameters themselves. (2) GAs search from a population rather than a single point.

(3) GAs use payoff (objective function) information, not derivatives or other auxiliary knowledge. (4) GAs use probabilistic transition rules, not deterministic rules. In addition it is found that GA is highly efficient means of searching a large solution space. Some computational details of GA application to optimize three-dimensional highway alignments (1) relevant to this study is described next. Data Format for Describing the Region of Interest A matrix format as shown in Figure 1, is employed to minimize the needed memory and carry important information for the entire region. The coordinates of the origin (bottom left corner) are labeled as O( xO , y O ) and the dimensions of each cell are Dx and Dy . We further denote x max and y max as the maximal X and Y coordinates of the study region. Decision Variables (Points of Intersections) In highway engineering processes, points of intersections ( Pi , see Figure 2) are used to initially locate alignments. Those points are then connected linearly to make tangent sections. Finally, appropriate curves are fitted to create a smooth and continuous alignment. Genetic algorithms adopted here exactly follow the above real engineering processes. Therefore, points of intersections ( Pi ) are the decision variables for alignment optimization and a set of points of intersections describes one specific highway alternative. In Figure 2, Ci and Ti mean points of curvature and points of tangency, respectively. For notational convenience, we further denote
T0 = P0 = S and C n +1 = Pn +1 = E as the start and end points of the alignment.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

Genetic Encoding of Alignment Alternatives Each point of intersection is determined by three decision variables, namely the X , Y , and Z coordinates (1, 4-5). For an alignment represented by n points of intersections, the encoded chromosome is composed of 3n genes. Thus, the chromosome is defined as:

= [1 , 2 , 3 ,......, 3n 2 , 3n 1 , 3n ] = x P1 , y P1 , z P1 ,......x Pn , y Pn , z Pn

(1)

where: = chromosome

i = the i th gene, for all i = 1,.......,3n

(x

Pi

, y Pi , z Pi = the coordinates of the i th point of intersection, for all i = 1,......., n

Genetic Operators The genetic operators employed for this study are problem-specific. Each operator is designed to work on the decoded points of intersections rather than individual genes. 1. Uniform Mutation Let = [1 , 2 , 3 ,......, 3n 2 , 3n 1 , 3n ] be the chromosome to be mutated at the encoded genes of the k th intersection point, where k = rd [1, n] , Then 3k 2 and 3k 1 will be replaced by:
3k 2 = rc [xO , x max ] 3k 1 = rc [ yO , y max ]

(2a) (2b)

2.

Straight Mutation Let = [1 , 2 , 3 ,......, 3n 2 , 3n 1 , 3n ] be the chromosome for mutation. We randomly

generate two independent discrete random numbers i and j , where i = rd [0, n + 1] ,

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

j = rd [0, n + 1] , i j , and i < j . Then the intermediate genes between the (3i ) th and (3 j 2) th

will be replaced by:


3l 2 = 3i 1 + (l i )

3 j 2 3 i 2
j i

for all l = i + 1,......, j 1

(3a)

j i

3.

Non-Uniform Mutation Let = [1 , 2 , 3 ,......, 3n 2 , 3n 1 , 3n ] be the chromosome to be mutated at the encoded

genes of the k th intersection point, where k = rd [1, n] . We first generate two random binary digit
rd [0,1] . Then the alleles of 3k 2 and 3k 1 in the resulting offspring = [1 , 2 , 3 ,..., 3k 2 , 3k 1 , 3k ,..., 3n 2 , 3n 1 , 3n ] are determined by the following rules:

(1)

If the first random digit rd [0,1] = 0 , then


3k 2 = 3k 2 f (t , 3k 2 xO )

If the first random digit rd [0,1] = 1 , then


3k 2 = 3k 2 + f (t , x max 3k 2 )

where: f is defined l = i + (l i )

(2)

If the second random digit rd [0,1] = 0 , then


3k 1 = 3k 1 f (t , 3k 1 y O )

TRB 2003 Annual Meeting CD-ROM

3l = 3i + (l i )

3 j 3i
j i

for all l = i + 1,......, j 1

( j i ) j i

3l 1 = 3i 1 + (l i )

3 j 1 3i 1

for all l = i + 1,......, j 1

(3b)

(3c)

(4a)

(4b)
for all l = i + 1,......, j 1

(5a)

Paper revised from original submittal.

Kim et al. If the second random digit rd [0,1] = 1 , then


3k 1 = 3k 1 + f (t , y max 3k 1 )

(5b)

4.

Whole Non-Uniform Mutation This operator applies the non-uniform operator to each point of intersection of a given

chromosome in a randomly generated sequence to change the entire configuration of the corresponding horizontal alignment. 5. Simple Crossover Let two parents i = i1 , i 2 , i 3 ,......, i ( 3n 2 ) , i ( 3n 1) , i ( 3n )
j = j1 , j 2 , j 3 ,......, j ( 3n 2 ) , j ( 3n 1) , j ( 3n )

and

be crossed after a randomly generated position

3k , where k = rd [1, n] . Then the resulting offspring are = i1 , i 2 , i 3 ,..., i ( 3k ) , j ( 3k +1) ..., j ( 3n 2) , j ( 3n 1) , j ( 3n ) i j = j1 , j 2 , j 3 ,..., j ( 3k ) , i (3k +1) ..., i ( 3n 2 ) , i (3n 1) , i ( 3n )

] ]

(6a) (6b)

6.

Two-point Crossover Let i = i1 , i 2 , i 3 ,......, i ( 3n 2 ) , i ( 3n 1) , i ( 3n )

and

j = j1 , j 2 , j 3 ,......, j ( 3n 2 ) , j ( 3n 1) , j ( 3n )

be the two parents to be crossed between two

randomly generated positions 3k and 3l , where k = rd [1, n] , l = rd [1, n], k l , and k < l . Then the resulting offspring are
= i1 , i 2 , i 3 ,..., i ( 3k ) , j ( 3k +1) ..., j ( 3l ) , i ( 3l +1) ,..., i ( 3n ) i

] ] ]
and

(7a) (7b)

j = j1 , j 2 , j 3 ,..., j ( 3k ) , i ( 3k +1) ..., i ( 3l ) , j ( 3l +1) ,..., j ( 3n )

7.

Arithmetic Crossover Given two parents i = i1 , i 2 , i 3 ,......, i ( 3n 2 ) , i ( 3n 1) , i ( 3n )

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

j = j1 , j 2 , j 3 ,......, j ( 3n 2 ) , j ( 3n 1) , j ( 3n ) , the arithmetic crossover reproduces two offspring

as follows:
= i + (1 ) j i j = j + (1 ) i

(8a) (8b)

where = rc [0,1] 8. Heuristic Crossover Let the two parents to be crossed by this operator be denoted by
i = i1 , i 2 , i 3 ,......, i ( 3n 2 ) , i ( 3n 1) , i ( 3n )

and

j = j1 , j 2 , j 3 ,......, j ( 3n 2 ) , j ( 3n 1) , j ( 3n ) , where we assume CT ( i ) CT ( j ) (i.e., i is

at least as good as j ). Then the operator generates a single offspring according to the following rule:
= ( i j ) + i

(9)

where = rc [0,1] Further details on genetic encoding and operators can be found in Jong et al. (4), and Jong and Schonfeld (5).

METHODOLOGY
When obtaining an alignment alternative through an optimization process, the expected outputs are three-dimensional coordinates of the alignment centerline. To describe highway alignments (or centerlines of highways), a parametric representation is useful (30-32). In the proposed method a smooth and continuous alignment is explored in a given solution space.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

Boldface capital letters will be used to denote vectors in space. Let P (u ) = [ x(u ), y (u ), z (u )]T be

a position vector along the alignment L , where u =

P (t ) dt P (t ) dt
0 0 1

and

P (u ) = ( x (u )) 2 + ( y (u )) 2 + ( z (u )) 2 . Basically, P is parameterized by u , which represents

the fraction of arc length traversed to that point. If L is an alignment connecting


S = [ x S , y S , z S ]T and E = [ x E , y E , z E ]T , then the position vector P (u ) must satisfy P (0) = S ,

and P (1) = E . P (u ) must also be continuous and continuously differentiable in the interval
u [0,1] .

Alignment Optimization Model Formulation Model formulation consists of two parts: (1) objective function and (2) constraints. The objective function is the total cost function having five main components (user cost ( CU ), rightof-way cost ( C R ), pavement cost ( C P ), earthwork cost ( C E ) and structure cost ( C S )) as shown in Equation (10).
x P1 , y P1 , z P1 ,....., x Pn , y Pn , z Pn

Minimize

CT = CU + C R + C P + C E + C S i = 1,....., n i = 1,....., n

(10) (10a) (10b)

subject to xO x Pi x max ,
y O y Pi y max ,

where ( xO , y O ) = the X , Y coordinates of the bottom-left corner of the study region (Fig. 1)
( x Pi , y Pi ) = the X , Y coordinates of points of intersections, Pi ( x max , y max ) = the X , Y coordinates of the top-right corner of the study region (Fig. 2)

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

The user cost consists of traveltime cost, vehicle operating cost, and accident cost (4-5, 33). The right-of-way cost consists of the land area taken by the alignment and damage to the properties (34). There are also many design and operational constraints to be met in alignment optimization. Among those, important factors can be addressed as follows: (1) Alignment Necessary Conditions Any point of the horizontal alignment ( Lxy ) of an alignment (L) should be a part of the set of tangent sections ( C t ) or circular curves ( C c ) or spiral curves ( C s ). (2) Horizontal Curvature Constraint (Minimum Radius Constraint) The degree of horizontal curvature should be less than the maximum allowable value,
H max , to provide safe turning radius. A common way to express this constraint is through the

minimum radius constraint for horizontal alignments.


Rmin = Vd 15(e + f s )
2

(11)

where Vd =design speed (mph)


e = superelevation f s = coefficient of side friction

(3) Gradient Constraint To avoid abrupt change over the vertical alignment of highways, the maximum allowable gradient constraint should be specified.
dz (u ) Gmax , u [0,1] dh(u )

(12)

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. where h(u ) = ( x (t )) 2 + ( y (t )) 2 dt


0 u

10

(4) Vertical Curvature Constraint As in the case of horizontal curvature constraint, parabolic curves curvature in vertical alignments should be less than the maximum allowable value, Vmax . This constraint can be expressed as the minimum length of the vertical curve, Lm (35-37). Crest Curve
Lm =

100 2hd + 2ho

2 A S d

if Lm > S d

(13)

Lm = 2 S d

100 2hd + 2ho A

, if Lm < S d

(14)

where Lm = minimum length of vertical curve (ft)


A = algebraic difference in grades (percent), g i g i 1 S d = sight distance (ft), S d (i ) = 3.67Vd + Vd2 30( f r + g i )

hd = height of driver's eye above roadway surface (ft) ho = height of object above roadway surface (ft)

Sag Curve
Lm =
2 A S d , 400 + 3.5S d

if Lm > S d (or A >

400 + 3.5S d ) Sd

(15)

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.
Lm = 2 S d 400 + 3.5S d , A

11 if Lm < S d (or A <


400 + 3.5Sd ) Sd

(16)

The vertical alignment constraints are considered in the model in the form of penalties (1-3). Additional constraints on structures (i.e., bridges and tunnels) need to be considered. For example, to select between fills and bridge construction, we need to know elevation differences between ground elevations and road elevations. This elevation difference also applies for constructing tunnels rather than cuts. Some constraints for intersection, overpass, underpass, and interchange construction should also be considered. Examples of such

constraints are: crossing angle constraint, vertical and lateral clearance constraints and land use, budget and regulations (6). A more detailed explanation of how vertical alignment constraints and other constraints are incorporated in the model is available in Jong (1) and Kim (3). An Alignment Optimization Example Using Genetic Algorithms This section illustrates how the alignment optimization using genetic algorithms works, and what kind of results it produces. Figure 3 shows an artificial study area with fairly complex topography that includes a two-lane highway from the center of North to South-East, three hills and a creek crossing from North-East edge to South. Darker cells represent higher elevations. The grid size is 200200 ft. Our plan is to build a two-lane highway connecting the start and end points (Fig. 3) while allowing the existing road to be re-optimized. The following values are used to compute the cost functions that make up the objective function: design speed (mph) 50; coefficient of side friction (decimal) 0.16; superelevation (decimal) 0.06; maximum allowable grade (%) 5; coefficient of forward rolling friction (decimal) 0.28; filling slope (decimal-tangent value) 0.4 (2.5:1); cutting slope (decimal-tangent value) 0.5; earth shrinkage factor 0.9; unit pavement cost ($/ft) 0.1; unit cost for diesel fuel ($/gallon) 0.85; unit cost for gasoline ($/gallon)

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

12

1.25; average accident cost ($/per accident) 20000; unit cutting cost ($/cub yard) 35; unit filling cost ($/cub yard) 20; unit transportation cost for moving earth from a borrow pit ($/cub yard) 2; unit transportation cost for moving earth to a landfill ($/cub yd) 3; analysis period (years) 30; interest rate (decimal) 0.06; annual average daily traffic 2000; traffic growth rate (decimal) 0.005. In the interest of the page limitations set by TRB it is not possible to give all the details on how these values are used in the model. Readers should refer to (1-5, 33). Although the model is designed to automatically select the best crossing type of the new alignment with the existing road, in this example it is assumed that users specify an intersection as the crossing type with the existing road. A desktop computer with 1 GHz CPU speed and 261 MB RAM is used to run the program. Figures 4 shows the optimized solution and other useful information. The figure shows three main window areas: (1) horizontal alignment, (2) vertical alignment and (3) generation number and best solution value. The best solution contains two bridges, two tunnels and an intersection crossing the existing road with approximately 70 degrees. Table 3 provides general information for the test run. Computation time took 4 minutes and 50 seconds for 500 generations. Since the existing road initiated an additional module for intersection evaluation, 4 minutes and 50 seconds are found to be relatively longer when considering other types of structures. For instance, 3 minutes and 24 seconds took for a grade separation and 3 minutes and 25 seconds consumed for an interchange. Please note that this is a relatively simpler example in which saving a few minutes of computing time may not be very significant; however, for larger problems with heterogeneous land use, especially when the model is connected to a GIS (2) saving in computing time assumes particular significance.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

13

Total costs are found to be 21.03 million for approximately 1.5 miles long alignment. It is also found that user costs account for 33% of the total costs. Structures costs including an intersection, two bridges and two tunnels are 28% while construction costs only account for 39%.

STEPWISE ALIGNMENT OPTIMIZATION


Now, our concern is to examine if the stepwise approach yields any improvement in computational efficiency and the quality of solution. In many optimization processes, subdividing large problems into suitable pieces can decrease the computation time and produce a better solution. This argument also applies to this study, since modeling intersections and other structures in alignment optimization repeatedly involves fine-tuning search steps for structures. Another issue for computational efficiency and search performance is the population size. Goldberg (29) has shown that the efficiency of a GA in reaching a global optimum instead of local ones largely depends on the population size. In our application, the population size for each generation is set proportionally to the number of decision variables (points of intersections, Pi s). For example, if three points of intersections are used for generating highway alignments, then the population size is set at 30 (= 3 10) while a population of 150 is used for 15 points of intersections. The artificial study area previously used is chosen for a stepwise alignment optimization and three scenarios shown in Table 4 are designed to check the search performance and computational efficiency. Scenarios 1 and 2 are devised for a one-step optimization while scenario 3 is for a two-step optimization. The results of scenarios 1 and 2 can be used for assessing the effects of the population size on computational time and the quality of solutions

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. while the result of scenario 3 can be compared to the results of both scenario 1 and 2 for checking how much improvement is found with a two-step optimization.

14

The crossing type with the existing road (Fig. 3) is again assumed to be intersection, to preserve comparison basis. The unit excavation cost is assumed to be $100 and the limiting value beyond which tunnels are considered rather than cuts is assumed to be 20 ft. Since the optimization processes using GA is stochastic, each program run shows different results. Therefore, several runs (we call it replications) need to be made to check (1, 4) the variance of results. Therefore, three replications are run for scenarios 1 and 2. Table 5 shows comparison between two scenarios and Figures 5 and 6 show the best solutions among three replications under scenarios 1 and 2. In scenario 1, two bridges, one tunnel and an intersection are found while scenario 2 shows one bridge, an intersection and no tunnels in the best solutions. Total costs of the best solutions for each scenario significantly decreased from $22.14 million to $17.29 million ($4.85 million, 21.9% improvement) while computation time for scenario 2 is 4.72 times longer for scenario 1. These results indicate various tradeoffs between cost and computational time. Please note that GA does not guarantee a global optimal solution rather it gives a near optimal solution. Also, for a problem such as ours it is possible to obtain significantly different solution values for slightly different alignments requiring bridge/tunnel constructions versus cut/fill. Therefore, caution should be exercised in interpreting the applications of the stepwise approach. To check computation time and the quality of solutions of the two-step optimization, the three points of intersections of scenario 1 are obtained after the one-step optimization and used to subdivide the whole alignment into four segments. Figure 7 shows the resulting segmentation. Since direct use of the points of intersection for the start or end points of each segment may not

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

15

insure the smoothness when connecting the solutions from each segment, a 150 ft gap is inserted between segments and its costs are averagely added to adjoining segments. It is notable that only segment 3 requires evaluation of intersections. Table 6 shows the results of two-step optimization under scenario 3. Total computation time for scenario 3 is 6 minutes and 34 seconds; an improvement of more than two hours over scenario 2 is noted. Also the computing time for scenario 3 is about 20% of that for scenario 1. Also, it is worthwhile to note that the computation times for segments 1, 2 and 4, which do not require intersection evaluation, are less than 20 seconds. More importantly, the overall costs of each segment are found to be $16.74 million (76% of one-step optimization solution of scenario 1). Thus, two-step optimization not only improves the computing efficiency but also finds a better solution. It is important to note here that one stage optimization is performed only to compare the computing time and solutions to the stepwise approach. Also, as noted earlier for more complex examples these improvements will be fairly significant. Figures 8, 9, 10 and 11 show the optimized solutions for each segment with two-step optimization. Significant changes compared to one-step solution (Fig. 5) has been noted in the solution of segment 4. For instance, the tunnel found in one-step solution has disappeared. Instead, only cuts and fills are observed in the optimized alignment which is shifted to the right, costing less than tunnel construction.

CONCLUSIONS
Highway alignment optimization is a complex problem. It has an infinite number of alternatives to be evaluated in a continuous search space. Moreover, cost functions are very difficult to formulate, are non-differentiable, noisy and implicit. Due to the nature of the problem faster and efficient search algorithms are needed rather than conventional methods.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

16

This study presented a stepwise highway alignment optimization procedure using genetic algorithms, one of the artificial intelligence (AI) techniques. The stepwise optimization is based on different population sizes and segmentation of study areas into suitable pieces. The proposed stepwise approach is implemented in an artificial test example, which indicates that substantial improvement in computing efficiency can be achieved with the stepwise approach. The approach also improves the solution (i.e., an economical alignment is obtained) compared to the traditional one stage approach. More test cases with larger problem size and additional GA scenarios are needed to be run to investigate the full potential of the stepwise approach. To subdivide a study area, it is recommended that a one-step optimization be run with a relatively small number of decision variables ( Pi s). Then the relevant Pi s location for subdivision should be selected based on: (1) the possibility for construction of structures and (2) the precision requirements. The lengths of segments may also differ depending on the precision requirements and need for savings in computing time. The method was not implemented on a real map using a GIS. Further research is necessary to examine how much improvement in computational efficiency and the quality of solutions can be achieved when the stepwise optimization is adopted for real application.

ACKNOWLEDGEMENTS
The authors wish to thank the four anonymous reviewers whose valuable comments enhanced the quality of the paper. This research has been partially performed by the Advanced Highway Research Center funded by the Korea Science and Engineering Foundation affiliated to the Korea Ministry of Science and Technology.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

17

REFERENCES
1. Jong, J.-C. Optimizing Highway Alignments with Genetic Algorithms, University of Maryland, College Park. Ph. D. dissertation, 1998. 2. Jha, M. K. A Geographic Information Systems-Based Model for Highway Design Optimization. University of Maryland, College Park. Ph. D. dissertation, 2000. 3. Kim, E. Modeling Intersections and Other Structures in Highway Alignment Optimization. University of Maryland, College Park. Ph. D. dissertation, 2001. 4. Jong, J.-C., Jha, M. K., and Schonfeld, P. Preliminary Highway Design with Genetic Algorithms and Geographic Information Systems. Computer Aided Civil and Infrastructure Engineering, 15(4), 2000, pp. 261-271. 5. Jong, J.-C. and Schonfeld, P. An Evolutionary Model for Simultaneously Optimizing 3Dimensional Highway Alignments. Transportation Research-B, in press. 6. OECD. Optimisation of Road Alignment by the Use of Computers. Organization of Economic Co-operation and Development, Paris, 1973. 7. Shaw, J. F. B. and B. E. Howard. Expressway Route Optimization by OCP, ASCE Journal of Transportation Engineering, Proceedings of the American Society of Civil Engineers, ASCE, Vol. 108, No. TE3, 1982, pp. 227-243. 8. Fwa, T. F. Highway Vertical Alignment Analysis by Dynamic Programming, Transportation Research Record 1239, TRB, National Research Council, Washington, D.C., 1989, pp. 1-9. 9. Nicholson, A.J., D.G. Elms, and A. Williman. A Variational Approach to Optimal Route Location. The Journal of the Institution of Highway Engineers, Vol. 23, March 1976, pp. 2225.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

18

10. Howard, B.E., Z. Brammick, and J.F.B. Shaw. Optimum Curvature Principle in Highway Routing. Journal of the Highway Division, ASCE, Vol. 94, No. HW1, 1968, pp. 61-82. 11. Thomson, N.R. and J.F. Sykes. Route Selection through a Dynamic Ice Field Using the Maximum Principle. Transportation Research-B, Vol. 22B, No. 5, May 1988, pp. 339-356.3. 12. Shaw, J.F.B. and B.E. Howard. Comparison of Two Integration Methods in Transportation Routing. Transportation Research Record 806, 1981, pp. 8-13. 13. Turner, A.K. and R.D. Miles. A Computer Assisted Method of Regional Route Location. Highway Research Record 348, 1971, pp. 1-15. 14. Athanassoulis, G.C. and V. Calogero. Optimal Location of a New Highway from A to B - A Computer Technique for Route Planning. PTRC Seminar Proceedings on Cost Models and Optimization in Highways (Session L9), London, 1973. 15. Parker, N.A. Rural Highway Route Corridor Selection. Transportation Planning and Technology, Vol. 3, 1977, pp. 247-256. 16. Trietsch, D. A Family of Methods for Preliminary Highway Alignment. Transportation Science, Vol. 21, No. 1, 1987a, pp. 17-25. 17. Trietsch, D. Comprehensive Design of Highway Networks. Transportation Science, Vol. 21, No. 1, 1987b, pp. 26-35. 18. Hogan, J.D. Experience with OPTLOC - Optimum Location of Highways by Computer. PTRC Seminar Proceedings on Cost Models and Optimization in Highways (Session L10), London, 1973. 19. Easa, S. M. Selection of Roadway Grades that Minimize Earthwork Cost using Linear Programming. Transportation Research-A, Vol. 22, No. 2, 1988, pp. 121-136.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

19

20. Puy Huarte, J. OPYGAR: Optimization and Automatic Design of Highway Profiles. PTRC Seminar Proceedings on Cost Models and Optimization in Highways, Session L13, 1973, London, UK. 21. Murchland, J. D. Methods of Vertical Profile Optimization for an Improvement to an Existing Road. PTRC Seminar Proceedings on Cost Models and Optimization in Highways, Session L12, 1973, London, UK. 22. Goh, C. J., Chew, E. P., and Fwa, T. F. Discrete and Continuous Model for Computation of Optimal Vertical Highway Alignment. Transportation Research-B, Vol. 22, No. 9, 1988, pp. 399-409. 23. Hayman, R. W. Optimization of Vertical Alignment for Highways through Mathematical Programming. Highway Research Record 306, 1970, pp. 1-9. 24. Robinson, R. Automatic Design of the Road Vertical Alignment. PTRC Seminar Proceedings on Cost Models and Optimization in Highways, Session L9, 1973, London, UK. 25. Chew, E. P., Goh, C. J., and Fwa, T. F. Simultaneous Optimization of Horizontal and Vertical Alignments for Highways. Transportation Research-B, Vol. 23, No. 5, 1989, pp. 315-329. 26. Wan, F. Y. M. Introduction to the Calculus of Variations and its Applications, 1995, Chapman & Hall, New York. 27. ReVelle, C. S., Whitlatch, E. E., and Wright, J. R. Civil and Environmental Systems Engineering, 1997, Prentice Hall, New Jersey. 28. Chapra, S. C. and Canale. R. P. Numerical Methods for Engineers, 1988, McGraw-Hall, Inc., New York.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. 29. Goldberg, D. E. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley Publishing Company, Inc., Massachusetts, 1989.

20

30. Mortenson, M. E. Geometric Modeling, 2nd Edition, John Wiley & Sons, Inc., New York, 1997. 31. Swokowski, Earl W. Calculus with Analytic Geometry, 2nd Edition, Prindle, Weber & Schmidt, Boston, Massachusetts, 1979. 32. Lovell, David J. Automated Calculation of Sight Distance from Horizontal Geometry, ASCE Journal of Transportation Engineering, Vol. 125, No. 4, July/August, 1999, pp. 297-304. 33. Jong, J.-C. and Schonfeld, P. Cost Functions for Optimizing Highway Alignments. Transportation Research Record, 1659, pp. 58-67. 34. Jha, M.K., Schonfeld, P. Geographic Information Systems-Based Analysis of Right-of-Way Cost for Highway Optimization. Transportation Research Record 1719, 2000, pp. 241-249. 35. AASHTO. A Policy on Geometric Design of Highways and Streets. American Association of State Highway and Transportation Officials, Washington, D. C., 2001. 36. Wright, P. H. Highway Engineering. John Wiley & Sons, Inc., New York, 1996. 37. Underwood, R. T. The Geometric Design of Roads. The Macmillan Company of Australia pty ltd, Australia, 1991.

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. List of Tables and Figures

21

Table 1 Studies on Highway Alignment Optimization Table 2 Defects of the Existing Highway Alignment Optimization Methods Table 3 General Information for the Test Run Table 4 Scenarios for a Stepwise Optimization Table 5 Comparison Between Scenarios 1 and 2 Table 6 Results for Each Segment of Scenario 3 Figure 1 An Example of Study Area for Alignment Optimization Figure 2 An Example of Points of Intersections, Tangency and Curvature Figure 3 Topography of the Artificial Study Area Figure 4 Optimized Solution for the Test Case Figure 5 The Best Solution among Three Replications for Scenario 1 Figure 6 The Best Solution among Three Replications for Scenario 2 Figure 7 Segmentation of the Scenario Solution 1 for 1 Two-Step Optimization Figure 8 Optimized Solution for Segment 1 under Scenario 3 Figure 9 Optimized Solution for Segment 2 under Scenario 3 Figure 10 Optimized Solution for Segment 3 under Scenario 3 Figure 11 Optimized Solution for Segment 4 under Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. TABLE 1 Studies on Highway Alignment Optimization Target for optimizing Types of approach Calculus of variations Network optimization Dynamic programming Genetic algorithms Enumeration Dynamic programming Linear programming Numerical research Genetic algorithms Dynamic programming Numerical research Two-Stage Optimization Genetic algorithms 7, 10-12, 26 6, 13-17 9, 18 1 19 8, 20-22 27-28 6, 8, 22-24 1 9, 18 25 15-16 1-3 References

22

Horizontal alignment

Vertical alignment

Horizontal and vertical alignment simultaneously

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. TABLE 2 Defects of the Existing Highway Alignment Optimization Methods Methods Calculus of variations Network optimization Defects Requires differentiable objective functions Not suitable for discontinuous factors Tendency to get trapped in local optima Outputs are not smooth Not for continuous search space Outputs are not smooth Not suitable for continuous search space Not applicable for implicit functions Requires independencies among subproblems Not suitable for continuous search space Inefficient Not suitable for non-linear cost functions Only covering limited number of points for gradient and curvature constraints Tendency to get trapped in local optima Complex modeling Difficulty in handling discontinuous cost items

23

Dynamic programming

Enumeration Linear programming

Numerical research

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. Table 3 General Information for the Test Run Generation no. at which best solution found 500 Total cost ($) 21.03 million Computation time 4 minutes 50 seconds Crossing Costs Type ($) Intersectio 1.49 n million No. of tunnels 2

24

No. of bridges 2

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. TABLE 4 Scenarios for a Stepwise Optimization Number of points of intersections ( Pi s) between start and end points for an one-stage optimization 3 15 Number of points of intersections ( Pi s) within each segment 3

25

Type of optimization

Scenarios

Subdivided highway segments

Population size

Number of generations

One-step optimization

Scenario 1 Scenario 2 Scenario 3

0 0

30 150 Population size for each segment 30

2000 2000 Number of generations for each segment 2000

Two-step optimization

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. TABLE 5 Comparison Between Scenarios 1 and 2


Type of optimization Scenarios Scenario 1 Scenario 2 Replications 1 2 3 1 2 3 Total costs (million) 22.29 22.14 22.43 20.78 17.29 18.07 Computation time (hour/minute/second) 0 / 32 / 34 0 / 31 / 14 0 / 31 / 18 2 / 39 / 8 2 / 27 / 38 2 / 35 / 20 Seeds 21088 20035 21276 21088 20035 21276

26

One-step optimization

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al. TABLE 6 Results for Each Segment of Scenario 3 Type of Scenarios optimization Generation at which Segments the optimized solution found Segment 1 1985 Segment 2 1905 Segment 3 Segment 4 Total N. A. N. A. 1998 1998 N. A. Total costs ($million) 4.07 5.01 3.69 3.97 16.74 Computation time 15 seconds 14 seconds 5 minutes 48 seconds 17 seconds 6 minutes 34 seconds

27

Two-step optimization

Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

28

y=ymax

Dy y=yO+2Dy Dx y=yO+Dy (xO,yO) x=xO+Dx

x=xmax

FIGURE 1 An Example of Study Area for Alignment Optimization

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

29

Cn+1=Pn+1=E T5 T2 T1 C3 C2 T3 C4 T4=C5

C1

T0=P0=S
FIGURE 2 An Example of Points of Intersections, Tangency and Curvature

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

30

End point

Creek

Start point

FIGURE 3 Topography of the Artificial Study Area

TRB 2003 Annual Meeting CD-ROM

elevation (ft)
500-600 400-500 300-400 200-300 100-200 0-100
Existing road

Paper revised from original submittal.

Kim et al.

31

Tunnel 2

Bridge 1

Tunnel 1

Bridge 2

FIGURE 4 Optimized Solution for the Test Case

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

32

Tunnel 1

Bridge 2 Bridge 1

FIGURE 5 The Best Solution among Three Replications for Scenario 1

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

33

Bridge 1

FIGURE 6 The Best Solution among Three Replications for Scenario 2

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

34

Segment 4

Segment 3

Segment 1

Segment 2

FIGURE 7 Segmentation of the Scenario Solution 1 for a Two-Step Optimization

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

35

Bridge

FIGURE 8 Optimized Solution for Segment 1 under Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

36

Bridge

FIGURE 9 Optimized Solution for Segment 2 under Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

37

FIGURE 10 Optimized Solution for Segment 3 under Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Kim et al.

38

FIGURE 11 Optimized Solution for Segment 4 under Scenario 3

TRB 2003 Annual Meeting CD-ROM

Paper revised from original submittal.

Вам также может понравиться