Using Simulated Annealing and Spatial Goal

Programming for Solving a Multi Site Land Use

Allocation Problem

Jeroen C.J.H. Aerts1, Marjan van Herwijnen1, and Theodor J. Stewart 2

1
Institute for Environmental Studies (IVM), Vrije Universiteit Amsterdam
De Boelelaan 1087, 1081 HV Amsterdam, The Netherlands
Tel: +31-20-4449 555, Fax: +31-20-4449 553
{Jeroen.Aerts Marjan.van.Herwijnen}@ivm.falw.vu.nl

2
Department of Statistical Sciences, University of Cape Town
Rondebosch 7701, South Africa
[email protected]

Abstract. Many resource allocation issues, such as land use- or irrigation

planning, require input from extensive spatial databases and involve complex
decision-making problems. Recent developments in this field focus on the
design of allocation plans that utilize mathematical optimization techniques.
These techniques, often referred to as multi criteria decision-making (MCDM)
techniques, run into numerical problems when faced with the high
dimensionality encountered in spatial applications. In this paper, it is
demonstrated how both Simulated annealing, a heuristic algorithm, and Goal
Programming techniques can be used to solve high-dimensional optimization
problems for multi-site land use allocation (MLUA) problems. The optimization
models both minimize development costs and maximize spatial compactness of
the allocated land use. The method is applied to a case study in The
Netherlands.

1 Introduction
The area of ’resource allocation’ deals with the spatial distribution of (natural-)
resources such as water or land. Land use allocation problems are complex as they
often involve multiple stakeholders with conflicting goals and objectives. Therefore,
much attention has been paid to solving land use allocation problems with multi
criteria decision-making techniques (MCDM). Recent research focused on combining
MCDM with a geographic information system (GIS). This appears to be a powerful
combination, since land use allocation problems both involve multiple objectives and
criteria as well as geographically dependent spatial attributes [2], [3], [8], [9] and
[10].

Both GIS and MCDM techniques are derived from relatively technical areas of
geography and operations research. Practical use of such techniques often requires a
thorough understanding, and non-technical decision makers may find using these
techniques difficult. However, combined GIS-MCDM techniques can be opera-

C.M. Fonseca et al. (Eds.): EMO 2003, LNCS 2632, pp. 448–463, 2003.
© Springer-Verlag Berlin Heidelberg 2003
 Using Simulated Annealing and Spatial Goal Programming 449

tionalized for non-technical users by integrating these techniques in a (spatial-)

decision support system (SDSS) that is dedicated to a user. A series of SDSSs were
built only using GIS based mapping and visualization tools to provide background
information on the planning area. Some SDSSs successfully combined multi criteria
analysis (MCA) techniques (a subset of MCDM) with GIS enabling the interactive
evaluation of spatial land use plans against a pre-defined set of criteria, e.g. [13], [22].
The combination MCA-GIS can be useful in an SDSS environment when both the
alternatives have been clearly defined and the total set of alternative solutions is
limited to say six and rarely ten [23].

A challenge lies in further developing combined MCDM and GIS techniques suitable
for implementation in an SDSS for land use allocation problems. For instance, in case
alternative solutions are not defined, optimization techniques (another subset of
MCDM techniques) can be combined with GIS to support the spatial design of land
use allocation plans. Optimization techniques with GIS may therefore be referred to
as spatial design techniques [2], [3], [26]. In this paper we examine the use of spatial -
GIS based- optimization techniques for multi site land use allocation problems
(MLUA). ‘Multi site’ refers to the problem of allocating more than one land use type
in an area. A crucial element in the model is to introduce a spatial compactness
objective. Spatial compactness objectives are used to address the problem of
allocating the same land use not only at lowest cost but also at maximum
compactness, e.g. [2], [10], [32], [33].

From the above, we arrive at the following objectives for this research:
(1) To develop a goal-programming model based on a reference point approach that
can solve an MLUA problem. The model will be solved using a simulated
annealing algorithm.
(2) To develop different spatial compactness objectives in order to provide a decision
maker with different options
(3) To test the model on the basis of its efficacy to encourage spatial compactness.

A case study in The Netherlands illustrates how the optimization model can be used
in a decision environment.

2 MLUA Model
2.1 The Basic Allocation Problem

As an example, consider a rectangular area to be allocated with land use. First, the
area is divided to a grid with N rows and M columns. Let there be K different land use
types. We now introduce a binary variable xijk which equals 1 when land use k is
assigned to cell (i,j) and equals 0 otherwise. Furthermore, development costs (Cijpk)
are involved with each land use type k in cell (i,j). These costs vary with location
because they may depend on specific cost attributes p (for p=1,…,P) of the area, such
as soil type, elevation and management costs.
450 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

The objective is to minimize costs associated with allocating land uses k (for
k=1,…,K) to a map u. Accordingly, the problem may be written as follows:

Minimize
K N M
f p (u ) = ∑∑∑ C ijpk xijk ∀ p = 1,K, P (1)
k =1 i =1 j =1

Subject to
K

∑x
k =1
ijk =1 ∀ i = 1, K , N , j = 1, K , M (2)

xijk ∈ {0,1}

Lk ≤ Ak ≤ U k (3)

where:
N M

∑∑ x
i =1 j =1
ijk = Ak ∀ k = 1,K, K (4)

and

∑A
k =1
k = N ⋅M (5)

Equation 2 specifies that one and only one land use must be assigned to each cell.
Because decision variable xijk must be either 0 or 1, the model is defined as an
integer program (IP). Equations 3, 4 and 5 bound the number of cells Ak allocated to a
certain land use type k between an upper and lower bound, expressed as Lk and Uk ,
respectively. This as compared to a stricter formulation:

N M

∑∑ x
i =1 j =1
ijk = N ⋅ M ⋅ Ok ∀ k = 1,K , K (6)

where a required fixed required area proportion for each land use type is represented
by parameter Ok [3]

2.2 Spatial Objectives

The additive objective function described in Section 2.1 can be expanded with a
second, spatial objective. This objective refers to spatial attributes as compactness or
 Using Simulated Annealing and Spatial Goal Programming 451

contiguity of land use of equal type. Note the difference between contiguity and
compactness in this respect (Fig. 1) [11]. Contiguity requires all cells of equal land
use to be connected (Fig. 1, middle). Compactness merely encourages cells of equal
land use to be allocated next to one another, but this may result in divided patches
(Fig. 1, right). In this paper, we will restrict ourselves to the compactness objective.

Fig. 1. Area with single land use (light gray) covering 52 cells. These cells are randomly placed
before optimization (left). The cells are allocated by optimizing contiguity (middle) and
compactness (right)

Spatial compactness objectives are, for instance, found in forestry research harvest
schedules, which deal with strict adjacency constraints e.g. [14], 19], [21]. Some
research in geographic information science have approached spatial compactness in
optimization modeling by rewarding cases where neighboring cells have equal land
use [2], [3], [4].

In order to accommodate flexibility for the user, it is proposed to develop a mix of

three spatial compactness measures, from which a user can choose or can use
combinations. The spatial compactness objectives merely address commonly used
compactness characteristics as size, perimeter and area of a cluster [8], [11]. We here
restrict ourselves to the following spatial compactness objectives:
œ Spatial objective 1: minimizing the number of clusters per land use type. Less
clusters of a certain land use type points to higher compactness and less
fragmentation. Hence, the ideal compactness value would be 1.
œ Spatial objective 2: maximizing the largest cluster relatively to the other clusters
identified under spatial objective 1. It is preferred having at least one large
compact cluster, rather than all clusters being compact but small. The ideal
compactness value would be again 1.
œ Spatial objective 3: minimizing the perimeter of a cluster. In order to transform
this measure size and scale independent, the perimeter is divided by the square
root of the cluster area. The ideal compactness value would be 4.

A simple algorithm, that counts the number of clusters, the area and perimeter of each
cluster (per land use type) has been developed by [29]. The calculation of the spatial
objectives is illustrated in Fig. 2. Here, the value for spatial objective 1 is 4, because 4
clusters of the same land use can be identified. The value for spatial objective 2 is
0.25, which can be calculated by dividing 1 (for identifying 1 largest cluster) by 4
(total number of clusters). The value for spatial objective 3 is calculated using the
following equation:
452 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

K R
H kr
∑∑
k =1 r =1 Lkr
(7)

where Hkr stands for the perimeter of an identified cluster r for land use k and Lkr
represents the area for each identified cluster r per land use k. The values for the
perimeters for cluster a, b, c and d in Fig. 2 are 20, 10, 12, and 22 respectively. The
values for the area of all clusters are 19, 6, 5 and 25. Hence by applying Equation 7,
the value for spatial objective function 3 becomes 4.59 + 4.08 + 5.37 + 4.4 = 18.44.

b
a

c
d

Fig. 2. Illustration of spatial compactness objectives for four clusters labeled a, b, c and d, in an
area. The values for the perimeters for clusters a, b, c and d are 20, 10, 12, and 22 respectively.
The values for the area are 19, 6, 5 and 25 for clusters a, b, c and d, respectively

In order to exclude single cells as a cluster in an optimal result, we may introduce an

extra constraint for a minimum cluster area value Sk for any cluster Lrk by:

Lrk ≥ Sk ( Fk ∈ Ak ) ∀ k = 1, K, K (8)

The total spatial compactness objective can be calculated using a weighted sum of the
above-described three spatial compactness measures. The difference of the spatial
objectives as opposed to the cost objective is the non-linearity in the spatial
formulations.

3 Goal Programming
3.1 Constraints and Criteria

The MLUA problem formulated above is clearly a multi objective problem, where
costs and compactness objectives have to be traded off against each other. In a
situation where decision makers know their goals but have difficulties with valuing or
weighting the relevant attributes involved in the multi criteria analysis, goal
programming is a commonly known technique to aid decision makers with their task.
We have chosen to use a generalized goal programming approach (reference point)
approach based on Wierzbicki [31]. For each objective, we define some goal or
reference point, say p for all cost attribute related goals and q for all spatial
objectives. The model should find a land use map u for which:
 Using Simulated Annealing and Spatial Goal Programming 453

f p (u ) ≤ γ p (9)

skq (u ) ≤ λkq (10)

Where f(u)p is the total value for all cost attributes p (p = 1 , …, P) and skq (x) the total
of spatial measures q (q=1,…,Q), which in this case is set to 3 (see section 2.2).

Wierzbicki [31] uses a ’scalarizing’ function, which measures under-achievement

relative to the goals, but placing the greatest weight on the least well satisfied goal.
Another commonly used scalarizing function can be found in the Tschebycheff
approach [27] where the goal is to minimize the sum of deviations relative to the
goals defined. We here use another scalarizing approach [28], which also minimizes
the sum of deviations but then relative to an ideal value. This approach can be defined
as follows:

Minimize:
ρ ρ
P  f p (u )− I p  K Q  s (u )− I 
∑ 
 γ − I
 +

∑∑ 

kq
λ − I
kq


p =1  p p  k =1 q =1  kq kq
(11)

Subject to: Equations (2), (3), (4) and (5)

In Equation 11, Ip is the best possible ideal value for each objective p if optimized on
its own, and is a suitably large power. A value of = 4 has been found to yield good
results [28]. Advantages of this approach is (1) to avoid the use of preferences or
weights, which are often difficult to interpret by users, and (2) the function is scale
free, which rules out the need for finding the worst performance levels to provide a
normalized scaling.

Constraint in Equation 3 is fuzzy. We may treat this constraint in the same manner as
goals, which is a useful contribution as this research aims at developing a goal
programming approach for MLUA problems. Treating fuzzy constraints as goals may
be implemented as follows. One should view the nominal constraint as an ideal, and
the goal then should minimize the deviation from this ideal:

ρ
 max{0; Lk − Ak } 
Lower bound objective:   (12)
 β k0 

ρ
Upper bound objective:  max{0; Ak − U k }  (13)
 
 β k0 
454 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

The scaling factor k is chosen such that a deviation of this magnitude corresponds to
the same level of satisfying as achieving the assumed goals for the costs and spatial
compactness objectives. It is proposed to handle these goals as extra penalty terms in
Equation 11.

4 Optimization Algorithms
4.1 Heuristic Algorithms

Since both an MLUA problem can be classified as a combinatorial optimization

problem, and our MLUA model is non-linear in its spatial objective formulations, we
need to find a heuristic optimization algorithm for solving the model. Note that
various researchers have developed linear MLUA models solved with an LP solver,
but ran against a limitation in the size of the spatial area that could be optimized [2],
[3], [4], [8], [9], [10]. Heuristic approaches, however, are robust, fast and capable of
solving large combinatorial problems, but they do not guarantee the optimal solution.
Applications of such algorithms for MLUA problems are simulated annealing, greedy
growing algorithms, genetic algorithms and tabu search [2], [3], [5], [6], [19], [21].
We here focus on using simulated annealing to solve the above described
optimization model.

4.2 Simulated Annealing

Kirkpatrick [15] introduced the concept of annealing in combinatorial optimization.

This concept is based on a strong analogy between combinatorial optimization and the
physical process of crystallization. This process has inspired Metropolis [20] to
propose a numerical optimization procedure known as the Metropolis algorithm (Fig.
3). This works as follows for our problem. The initial situation is the current land use
map u. The associated total development cost is denoted by f(0). Following the flow
diagram of Fig. 3, we now swap the land use of a randomly chosen cell into another
randomly chosen land use. The new land use must be both another land use as
compared by the land use in the current situation and one that is permitted by a
transition matrix (Table 2). This yields a new situation, with new development costs
f(1). Whether we accept the change from state 0 to state 1 depends on the difference
in costs f(1)-f(0). Once this is decided we repeat the swapping procedure, decide
whether the change is accepted, generate a new swap, and so on. Whenever the costs
f(1) are smaller than the costs f(0), the cell change is accepted. When f(1)>f(0), costs
are accepted with a certain probability following the Metropolis criterion Equation
(14) (see e.g. [1], [25]). This is achieved by comparing the value of the Metropolis
criterion with a random number drawn from a uniform [0,1) distribution (Fig. 3).

 f (0) − f (1) 
P( accept change) = exp  (14)
 s0 
where s0 is a control or freezing parameter.
 Using Simulated Annealing and Spatial Goal Programming 455

For Lm iterations do:

(Initial-) start situation

costs = f(0)

Land use change for random cell

New Situation
costs = f(1)

If f(0)<f(1) If f(1)>f(0)

Then
if exp ((f(0) – f(1) / s0) > random[0,1)

Then Else
f(0) = f(1) f(0) = f(0)

if number of iterations = Lm
Then:
m = m + 1, Calculate_Control (s0)

Fig. 3. Flow diagram of the simulated annealing algorithm

A crucial element of the procedure is the gradual decrease of the freezing parameter si
[17]. Usually, this is done using a constant multiplication factor:

si +1 = r ⋅ si (15)

where 0<r<1. Examples of simulated annealing studies for spatial optimization are
[5], [7], [19], [30].

5 Case Study Jisperveld

5.1 Introduction

Jisperveld is the largest connected brackish fen-meadow area of Western Europe. It is

situated in the Northwest of the Netherlands and measures about 2000 ha. The area is
well known by its high natural values represented by a selection of rare bird species
and wetland vegetation types. The Jisperveld area is subject to a debate on how to
456 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

both plan and manage the area in the future. It appears that the governmental planning
policy for land use is changing from predominantly agriculture to a combined
agriculture, nature area. Therefore, planners need to develop an optimal location for
two new land use types ‘extensive agriculture’ and ‘water (limited access)’, both of
which are not yet present. In order to apply our goal programming approach, we’ve
simplified the current land use map of the area.

Both the number of land use types present in the area and the resolution of the grid
cells was reduced (Fig. 4)

Intensive Agriculture
Extensive Agriculture
Residence
Industry
Recreation (day)
Recreation (day/night)
Wet natural area
Water (recreational use)
Water (Limited use)

Fig. 4. Location of the Jisperveld area in The Netherlands indicated with the black dot (left) and
the simplified current land use map of the Jisperveld, measuring 20 x 20 cells (right)

Experts made clear that closed patches of land use, like large closed areas of extensive
agriculture and water (Limited use), represent a higher natural value than fragmented
areas. Also, less fragmented areas have an increased potential for recreational
activities. A simplification was to ignore the spatial requirement for creating
corridors. This requirement would involve an adjustment of the compactness
objective.

5.2 Objectives

Following our model formulations, the six objectives distinguished in the Jisperveld
can be categorized in additive cost attributes and objectives that relate to maximizing
compactness of land use of the same type. The objectives are to maximize natural and
recreational values, minimize cost for changing land use and to maximize
compactness following the three describes spatial objectives. Note that both natural
and recreational values can be seen as costs that contribute positively to the overall
objective function. The values related to objectives 1 and 2 are defined as Cijpk,
summed over all cells in the area, for each land use type k (for k=1,…, K). Values are
defined by value maps, which either have a uniform value (a number in Table 1) or a
variable value (‘map’ in Table 1). All values are scaled between 1 and 10.

Furthermore, transition costs for changing current land use kc into future land use kf
are presented in Table 2. Management costs for maintaining certain land use types are
not considered in this model, but can be easily integrated following the same
approach as with objectives 1 and 2, but then by minimizing management costs.
 Using Simulated Annealing and Spatial Goal Programming 457

Table 1. Nature values and recreation values for each land use type

Land use type (k) Nature value Recreation value Range:

Range: [1, 10] [1, 10]
1. Intensive agriculture 5 6
2. Extensive agriculture Map Map
3. Residence 2 4
4. Industry 1 1
5. Recreation (day trippers) 6 Map
6. Recreation (overnight) 6 Map
7. Wet natural area Map 7
8. Water (recreational use) 6 Map
9. Water (limited access) Map 3

Table 2. Transition matrix, showing costs [euro/cell] to change current land use kc into future
land use kf

Current land use kc

Int. Ext. Resi- In- Recr. Recr. Wet Wa- Wa-
agri. agri. dence dus- (day (over natu- ter ter
Future Land use type (kf) try tripp. night re (recr) (limit
) ) )
1. Intensive agriculture 0 1000 1000 500 – 7000 – – –
0
2. Extensive agriculture – 0 – – – – – – –
3. Residence – – 0 – – – – – –
4. Industry – – – 0 – – – – –
5. Recreation (day trippers) – – 9000 – 0 5000 – – –
6. Recreation (overnight) – – – – – 0 – – –
7. Wet natural area – – – – – – 0 – –
8. Water (recreational use) – – – – – – – 0 1000
9. Water (limited access) – – – – – – – – 0

The three spatial objectives are related to the extent to which the different land uses
are connected or fragmented across the region. For this, the number of clusters for
each land use type, as well as the area and perimeter of each cluster is measured with
a simple counting algorithm (see for details [29]).

5.3 Constraints

The values for the constraints are listed in Table 3. They refer to Equations 8, 12 and
13. Land use type ‘Intensive agriculture’ covers 157 grid cells but is limited to 130
grid cells in the new design. Extensive agriculture, conversely, does not occur in the
original map but has to cover at least 27 grid cells in the new design. Setting a
minimum cluster size assumes that it is not realistic to have a connected area that is
smaller than the minimum cluster size. Note, for example, that the minimum cluster
size for ‘extensive agriculture’ is smaller than for ‘intensive agriculture’.
458 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

Areas can be fixed to a certain land use type, by indicating areas that are not allowed
to change. These constraints can be set through a map holding the fixed land use
types. Fig. 5 shows the map with fixed land use types for this case study.

Table 3. Various spatial constraints for each land use type k

Land use type k Lower bound Upper bound Current Minimum

cluster size
Intensive agriculture 100 130 157 4
Extensive agriculture 27 57 0 3
Residence 28 35 28 3
Industry 2 4 7 2
Recreation (day trippers) 3 10 6 3
Recreation (overnight) 1 5 1 1
Wet natural area 4 20 8 3
Water (recreational use) 150 193 193 4
Water (limited access) 0 43 0 4

Intensive Agriculture
Extensive Agriculture
Residence
Industry
Recreation (day)
Recreation (day/night)
Wet natural area
Water (recreational use)
Water (Limited use)

Fig. 5. Land use map showing the areas with fixed land use types

6 Results
6.1 Initial Conditions

Three model runs are selected, each having different parameter settings, in order to
evaluate the model on its efficacy for generating compact patches of land-use. The
initial simulated annealing values are the same for all runs. The start value of the
freezing parameter was determined with a trial run, following [30] and [3]. The
freezing parameter was chosen such that within 500 trial iterations, 80% of all
calculated costs were greater than the original situation. The decrease parameter r was
set to 0.85 and the iteration length L per temperature stage was set to 1,000. Both r
and L are kept constant across all four runs. About ten runs using exactly the same
parameter settings were executed to test the consistence of the model. Three of those
runs are presented in Fig. 6. Although some runs start with considerably higher costs
as compared to other runs, all runs eventually arrive within approximately 38,000
iterations at a scalarizing value of about 2000.
 Using Simulated Annealing and Spatial Goal Programming 459

6.2 Spatial Designs

Run 1 evaluates compactness according to a standard parameter setting where cost

objectives and spatial compactness objectives are equally preferred. Fig. 7 shows
different stages in the iteration process using the standard parameter set. The map at
the far left shows the initial – random – situation.

Iterations against costs

1000000

800000
Run 1
Costs

600000

Run 2
400000

200000

Run 3
0
11800
14500
17200
19900
22600
25300
28000
30700
33400
36100
38800
41500
44200
46900
49600
1000
3700
6400
9100

Iterations

Fig. 6. Costs calculated by using the scalarizing function in Equation 11, against the total
number of iterations by the simulated annealing algorithm

The map to the far right shows the final situation, achieved after a total of 80,000
iterations. At this stage, the objective function could no longer be improved and the
iteration was terminated

Fig. 7. Different stages in the optimization process. Far left: initial situation; center left: after
5,000 iteration stages; center right: after 20,000 iteration stages; far right: final situation after
80,000 iteration stages

Within the Run 2, compactness is stimulated by increasing the weights for all spatial
objectives. The weights on the spatial objectives are set to twice the value for the
weights on the cost objectives. The result is presented in Fig. 8, left. Table 4 shows
460 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

indeed that the spatial objective values of Run 2 improve as compared to Run 1, but at
the cost of lower natural and recreational values and higher land use change costs.

For Run 3, an increased weight is set to allocate one specific land use, in this case
‘Water limited use’. It is shown that indeed a large compact cluster is formed in the
middle of the area. This is probably the least expensive area to allocate a compact
cluster of this land use type (Fig. 8, right).

Intensive Agriculture
Extensive Agriculture
Residence
Industry
Recreation (day)
Recreation (day/night)
W et natural area
W ater (recreational use)
W ater (Limited use)

Fig. 8. Results for runs 2 (left) and 3 (right)

Table 4. Objectives values for the three test runs

Nat Value Rec. Val. Change land Min Largest Min

use #clusters cluster perimeter
Run 1 1917 2389 175500 42 4.55 51.13
Run 2 1878 2258 179000 40 4.87 50.66
Run 3 1919 2412 198500 43 4.92 53.78

7 Conclusions and SDSS Implementation

The main goal of this paper was to investigate whether goal programming (GP)
combined with simulated annealing, is an attractive alternative for designing spatial
resource allocation alternatives.

We have developed a general GP approach to solve a MLUA problem. It is thought

GP is a well-known approach in a situation where decision makers know their goals
but have difficulties with valuing or weighting the relevant attributes involved in the
multi criteria analysis. We here successfully implemented a scalarizing approach
developed by [28], which minimizes the sum of deviations relative to an ideal value,
this as opposed to the Tschebycheff goal programming approach where the goal is to
minimize the sum of deviations relative to defined goals.

Three spatial compactness objectives have been developed, based on commonly used
compactness characteristics that address size, perimeter and area of a cluster of the
same land use. These objectives have been made operational within the above-
mentioned scalarizing function. Then model has been solved using the simulated
 Using Simulated Annealing and Spatial Goal Programming 461

annealing algorithm that has been used in similar studies by e.g. [3]. The efficacy of
the model for generating compact land use designs was evaluated in four model runs,
all having different parameter settings. It appeared that the model indeed generated
compact patches of land uses, which increased in size when the weights on the
compactness objectives was increased.

The case study in The Netherlands, to which the model was applied, clearly shows the
potential of the approach in a decision support setting. More attention is needed,
however, to further develop the GP approach in an SDSS, accessible for non-technical
users. Therefore, an important feature of such an SDSS will be to interactively specify
goal levels of the development cost and spatial attributes. With respect to the spatial
attributes, however, this is more difficult, and goal setting in the SDSS should be done
by:
œ pre define the worst and ideal levels of the spatial objectives
œ allowing the user to interactively select his preference for one of the three spatial
objectives
œ allowing the user to select spatial goals as a proportion of the distance between
worst and ideal values. E.g. by using a slide bar, which allows selecting values
between worst and ideal.

Other applications of the model approach can be found in forestry services and water
allocation issues. Further research on larger, more realistic maps is recommended as
well as to assess the possibilities to solve the model with other heuristic methods.

References
1. Aarts, E., Korst, J.: Simulated Annealing and Boltzman Machines. A Stochastic
Approach to Combinatorial Optimisation and Neural Computing, J. Wiley, New York
(1989)
2. Aerts, J.C.J.H.: Spatial Decision Support for Resource Allocation. Integration of
Optimization, Uncertainty analysis and Visualization techniques. PhD thesis, University
of Amsterdam, Thela Thesis publishers, Amsterdam (2002)
3. Aerts, J.C.J.H. and Heuvelink, G.B.M.; Using Simulating Annealing for Resource
Allocation. International Journal of Geographic Information Science 16 (2002) 571–587
4. Aerts, J.C.J.H., Eisinger, E, Heuvelink G.B.M, Stewart T.J.: Using Linear Integer
Programming for Multi Site Land Use Allocation. Geographical Analysis. (Accepted
Aug. 2002)
5. Boston, K. and Bettinger, P.: An Analysis of Monte Carlo Integer Programming,
Simulated Annealing, and Tabu Search Heuristics for Solving Spatial Harvest Scheduling
Problems. Forest Science 45 (1999) 292–301
6. Brookes, C. J.: A parameterized region-growing program for site allocation on raster
suitability maps. International Journal of Geographical Information Science 11 (1997)
375–396
7. Church, R.L., Stoms D. M., Davis, F.W.: Reserve Selection as a Maximal Covering
Allocation Problem. Biological conservation 76 (1996) 105–112
8. Cova, T. J.: A General Framework For Optimal Site Search. PhD thesis, University of
California, Santa Barbara (1999)
462 J.C.J.H. Aerts, M. van Herwijnen, and T.J. Stewart

9. Cova, T.J. and Church, R.L.: Contiguity Constraints for Single-Region Site Search
Problems. Geographical Analysis 32 (2000) 306–329
10. Cova, T.J. and Church R.L. Exploratory spatial optimization in site search: a
neighborhood operator approach. Computer, Environment and Urban Systems, 24 (2000)
401–419
11. Diamond, J. T. and Wright, J.R.: Efficient Land Allocation. Journal of Urban Planning
and Development, 115 (1989) 81–96
12. Greenberg, H.J.: Mathematical Programming Glossary.
http://www.cudenver.edu/~hgreenbe/glossary/glossary.html. Last checked, Jan. 2002.
13. Janssen, R.: Multi objective decision support for environmental problems. PhD thesis,
Free University, Amsterdam (1991)
14. Jones, G.J., Meneghin, B.J., Kirby, M.W.: Formulating Adjacency Constraints in Linear
Optimisation Models for Scheduling Projects in Tactical Planning. Forest Science, 37
(1991) 1283–1297.
15. Kirkpatrick, S., Gelatt, C.D., Vecchi, M.P.: Optimisation by Simulated Annealing,
Science 220 (1983) 671–680.
16. Kurttila, M.: The spatial structure of forests in the optimization calculations of forest
planning – a landscape ecological perspective. Forest Ecology and Management 142
(2002) 129–142
17. Laarhoven van P.J.M.: Theoretical and Computational Aspects of Simulated Annealing.
PHD Thesis, Erasmus University Rotterdam (1987).
18. Levine, B.: Presentation for EE652, April15th, 1999.
http://microsys6.engr.utk.edu/~levine/EE652/Slides/tsld001.htm. Last checked, May.
2002.
19. Lockwood, C. and Moore, T.: Harvest scheduling with spatial constraints: a Simulated
Annealing approach. Can. J. of For. Res. 23 (1993) 468 – 478
20. Metropolis, N., Rosenbluth, A., Rosenbluth, M., Teller, A., and Teller, E.: Equation of
state calculations by fast computing machines. Journal of Chemical Physics 21 (1953)
1087–1092.
21. Murray, A.T. and Church, R.L.: Measuring the efficacy of adjacency constraint structure
in forest planning models. Canadian Journal of Forest Resources, 25 (1995)1416–1424.
22. Pereira, J.M.C., Duckstein, L.: A Multiple criteria decision making approach to GIS
based land suitability evaluation, International Journal of Geographical Information
Systems, vol. 7 (1993) 407–424.
23. Ridgley, M.R., Heil, G.W.: Multi Criterion planning of protected-area buffer zones. In: E.
Beinat and P. Nijkamp (eds), Multi-criteria Evaluation in Land-Use Management, Kluwer
(1998).
24. Ridgley, M. Penn, D.C., Tran, L.: Multicriterion Decision Support for a conflict over
Stream Diversion and Land-Water Reallocation in Hawaii. Applied Mathematics and
Computation 83 (1997) 153–172
25. Rogowski, A.S., Engman, E.T.: Using a SAR Image and a Decision Support System to
Model Spatial Distribution of Soil Water in a GIS Framework. Integrating GIS and
Environmental Modeling. National Center for Geographic Information and Analysis
(NCGIA), Santa Fe (1996.)
26. Seppelt, R., Voinov A.: Optimization methodology for land use patterns using spatially
explicit landscape models. Ecological modeling 151 (2002) 125–142.
27. Steuer, R.: The Tchebycheff procedure of interactive multiple objective programming. In:
Karpak, B., Zionts, S (eds.) Multiple Objective Decision Making and Risk Analysis
Using Micro Computers. Springe-Verlag, Berlin (1986).
28. Stewart, T.J.: A multi criteria decision support system for R&D project selection. Journal
of the operational Research Society 42 (1991) 1369–1389.
 Using Simulated Annealing and Spatial Goal Programming 463

29. Stewart, T.J., Herwijnen, M. van, Janssen, J. A Genetic Algorithm Approach to Multi
Objective Land use Planning. In prep. (2002)
30. Sundermann, E.: PET Image Reconstruction Using Simulated Annealing, Proceedings of
the SPIE Medical Imaging 1995 Conference, Image Processing (1995) 378–386
31. Wierzbicki, A.P.: Reference point approaches. In: T. Gal, T.J. Stewart and T. Hanne
(eds.): Multi Criteria Decision Making: Advances in MCDM models, Algorithms, Theory
and Aplications. Kluwer Academic Publishers, Boston (1999)
32. Williams, J.C. and Revelle, C.S.; Reserve assemblage of critical areas: A zero-one
programming approach. European Journal of Operational Research 104 (1998) 497–509.
33. Wright, J., Revelle, C.S., Cohon, J.: A Multi objective Integer Programming model for
the land acquisition problem. Regional Science and Urban Economics 12 (1983) 31–53.