You are on page 1of 10

Industrial Plant Pipe-Route Optimisation

with Genetic Algorithms


Dae Gyu Kim1 , David Corne2
1 School of Cognitive and Computing Sciences,
University of Sussex at Brighton, Falmer, Brighton, BN1 9QH, UK
dgkim@cogs.susx.ac.uk
2 Parallel, Emergent, and Distributed Architectures Laboratory (PEDAL),
Department of Computer Science, University of Reading,
Whiteknights, Reading RG6 2AY, UK
D.W.Corne@reading.ac.uk

Abstract. The pipe-route design problem for heavy industrial plant

concerns minimising pipe material cost while satisfying constraints on


required interconnections and obstacle avoidance. This process is invariably done by human experts, but modern stochastic iterative search techniques allow the opportunity to automate this process. This study explores the possibility of automated industrial pipe-route design on three
test problems, using stochastic hillclimbing, simulated annealing, and genetic algorithms. The representation strategy is explained and discussed,
and results are presented which show great promise for genetic algorithms
in particular in this application area.

1 Introduction
Hard pipe route design problems arise when equipment (such as pumps, boilers,
and so on) need to be multiply interconnected by uid or gas-carrying pipes in
such a way as to satisfy various constraints and objectives, such as to avoid obstacles, avoid undue pressure variations, enable easy maintenance, and minimise
material cost, while of course satisfying the required interconnections. Pipe route
design is one of the most important steps involved in the design of large-scale
engineering plant such as ships, power plants, chemical plants, and so on. In
practice, the process mainly relies on human experts and their associated experience, since the range and complexity of the constraints tends to dissuade
attempts at automation.
In most industrial design practice, the locations of the various pieces of equipment involved are determined in advance. Given these locations, a Piping and
Instrumentation Diagram (P&ID) is prepared, which speci es the connections
which must be made. For example, gure 1, in combination with a list of required
connections (such as E1-nozzle1 connects to E2-nozzle7, E4-nozzle5 connects to
E4nozzle7, and so on . . . ), constitutes a P&ID.
A designer is then given the job of laying out the pipe routes in accordance
with the P&ID, while meeting several other constraints and objectives as mentioned above. Designers' skill and experience tends to lead to routing designs

10
9

O15

O20

E3
7

CONTROL ROOM

7
6
9

8 7
E1

2
1

O12

6 5
E5
4 O16

PUMP

TANK

8
9
O11

O17

O18

10

11

1 2
O2

O1

3
O3
9

O4
7

O6
O13

10
E4

TANK

11
1

O14
O19

O7
6

O8
4

7 E2
8
RESERVOIR
9
1
2
3

O9

O10

O5

10

Fig. 1. A Piping & Instrumentation Diagram


which generally work well, but the greatly time-consuming nature of the design
process [1] makes it practically impossible to explore alternatives or seriously
attempt to optimise the design. Modern methods for tackling awkward problems, such as genetic algorithms (GAs) and simulated annealing (SA), introduce
the opportunity to greatly speed up and perhaps improve design practice and
quality in this area. The work reported here hence explored the potential for
GAs and other stochastic methods on this problem.
Many routing problems have been addressed with GAs . For example, [2] use
a GA to optimise the routing between chips on multi-chip modules, while [3]
describe three forms of pipe-network optimisation problem: pipe-sizing, layout
design, and pressure-regulation. The problem described here, however, di ers
in certain key respects from others we have so far found in the literature. For
example, pipe-layout problems studied tend to concern nding the best way to
arrange connections between multiple demand and supply stations, however in
our case the interconnections are all pre-speci ed and the problem is instead to
minimise overall pipe-length taking obstacles into account. Connections which
need to be made are also pre-speci ed in electrical routing examples such as
[2], however, unlike in our case, such cases also typically involve designing the
relative layout of the items being interconnected, while having no analogue of
the potential to merge connections (see below).

Overview The remainder of this paper is set out as follows. Section 2 describes

the pipe routing design problem in more detail, and sets out the chromosome

representation used. Test problems, experiments, and results are then presented
in section 3, and then a concluding discussion appears in section 4.

2 Representing Pipe Routes


The problem we address here can be partly (but see later) viewed as optimising separately each of a collection of individual pipe connections. The essential
requirement for such a connection is that it be made in such a way as to avoid
obstacles. Taking pipe cost and manufacturing considerations into account, we
can be justi ed in considering only rectilinear connections, and we therefore can
conveniently view the problem in terms of choices of Steiner points [4]. Figure 2
illustrates this; we wish to make a rectilinear connection between the pump and
the storage tank, that is, between points A and B. Such a connection can take
one of two `basic routes', going either via Steiner point 0, or Steiner point 1.

Pump
Steiner Point 1
A
Tank
Steiner
Point 0
B

Fig. 2. Illustrating Steiner Points


In order to avoid obstacles, however, the basic routes must be diverted somehow. There are essentially two possible diversions, termed `escape graphs' [5],
around each obstacle. This is illustrated in gure 3. Hence, a single basic route
(ie: a route involving a particular choice of Steiner point) gives rise to 2k di erent
possible routes when there are obstacles in the way.
Associated with every connection which has to be made, then, is a collection
of potential routes derived from the two (in two dimensions) possible Steiner
points, and the appropriate obstacles in the P&ID. Figure 4 illustrates a general
case, in which the connection between `Start' and `End' must be made around
three obstacles (left hand side of the gure), giving rise to 6 possible routes as
detailed in the right hand side of the gure. As can be seen, Routes A and B are
the two that arise from the rst basic route (involving Steiner point 0) and the
k

Obstacle 4

Route 1_1

Route 2_2

Obstacle 9
Start
Route 2_1
Route 1_2
Route 3_2

Route 3_1

Obstacle 1
End

Fig. 3. The construction of possible routes around obstacles


single obstacle. The two obstacles obstructing the second basic route (involving
Steiner point 1) give rise to the four potential routes C,D,E and F.
Route F
Start

Route B
Steiner 1

Route E
Route D

Route A
Route C

Possible routes

Steiner 0

End

Fig. 4. Examples of possible routes around obstacles


Given the complete set of connections which need to be made and the locations of all equipment and obstacles { hence, given the P&ID { our strategy
is to preprocess the information to produce the choice of all possible routes for
each individual connection (Eg: as illustrated for the single example connection
in gure 4). A simple chromosome representation strategy is then to provide a
gene for each connection, alleles of each gene indexing the set of possible routes
for that particular connection.
Following preliminary experimentation on small test cases, the representation strategy we chose as most promising for the main work reported here was
one in which each gene (connection) had only two alleles, 0 or 1, representing respectively the two choices of basic route (eg: either via Steiner point 0 or Steiner

point 1). Interpretation of the chromosome then automatically determined and


used the minimal length route around obstacles along the chosen basic route
(choosing deterministically between ties). We were hence able to represent 2D
rectilinear pipe-network optimisation as a binary chromosome problem.

Evaluation

The rst step in interpreting a chromosome simply uses preprocessed information


to determine which potential route is represented in the chromosome for each
gene. That is, if gene displays allele 0 (1), then the route to be taken for the th
connection is the one via Steiner point 0 (1) which, taking obstacle avoidance
into account, has minimal length.
Three operations are now applied in order to interpret the chromosome as a
complete pipe route. These are: Merging, Simpli cation, and Equivalent-length
calculation. Each is discussed brie y next.
i

Merging Total pipe length is calculated as the sum of each pipe connection
minus the amount of duplicate pipe length. Duplicate pipe length is the total
length of spatially coincident pipe sections that can in practice be run as a
single pipe. This typically occurs when pipe routes for di erent connections represented in the chromosome coincide in direction of ow in addition to location.
For example, in Figure 5, a combination of routes B and C is preferred to other
combinations, since this will reduce the total pipe length most. Note that the
fact that pipes can be merged in this way is what stops the pipe-route design
problem addressed here from being able to be trivially decomposed into optimising individual connections. Such pipe-merging becomes more and more necessary
with larger installations.

B
Co

nn
Co

ec

tio

ion

n2

ect

nn

Equipment 1

Equipment 2
D

Equipment 3

Fig. 5. Merging duplicate pipe routes

Simpli cation At this stage, it may be possible to make several simpli cations
in the pipe route so far, as a result of opportunities which may have arisen mainly
from obstacle avoidance. Certain common example candidates for simpli cation
are shown in gure 6. The pipe-route is checked for such opportunities and the
simpli cations are made.
1

==>
3

4
5

Channel reduction
6

==>

2
7

Bay reduction

==>

Point reduction

Fig. 6. Cases for pipe-route simpli cation

Equivalent-Length Calculation Finally, each corner in the pipe network is


converted into its equivalent length. This is a way of taking corners into consideration, whereby each corner is converted into a straight pipe with a certain length
which gives the same uid pressure drop as the corner [6], and the amount is
added to the total pipe length of the system.
Remarks Mainly owing to the merging and simpli cation processes, pipe-route

interpretation gives rise to a complex and time-consuming evaluation function.


To avoid undue complexity, the interpretation process as described above is
therefore `incomplete' in certain respects { for example, simpli cation may lead
to new opportunities for merging, but there is no extra merging step following
the simpli cation step. Also, simpli cation should ideally be iterated until no
more opportunities exist, but this would be particularly time-consuming. Also,
of course, it is built in to the interpretation function that, for each of the two (or
occasionaly one) Steiner points relevant to a connection, only one of the several
choices for potential routes around obstacles is allowed to be chosen. Allowing
more possible routes to be encoded may of course lead to better opportunity for
merging and simpli cation, leading to better overall solutions. This is yet to be
investigated fully, but preliminary experiments suggested that the representation
employed here struck a generally more robust balance between time complexity
and solution quality.

3 Experiments
We investigated the use of GAs [7], SA [8], and stochastic hillclimbing (SH) on
each of the test problems. In the following subsections we brie y overview the
experimental setup, and summarise the results.

Test Problems Three moderately sized test problems were developed, consisting of 20, 30, and 40 connections respectively. Figure 1 has already shown
the layout for the 20-connection problem, and 7 shows the optimal pipe route,
which we know since the solution space in this case was small enough to allow an
exhaustive search. Full details of all test problems are freely available from the
authors, and a web-site is currently being prepared to disseminate these details
more accessibly.

Fig. 7. The global minimum case with the 20-connection problem


Exhaustive search of the 20-connections problem results in the landscape
displayed in gure 8. The horizontal axis represents the numerical value of the
binary chromosome (in millions), from chromosome 0000 on the left-hand side
to chromosome 1111 on the far right, while the vertical axis displays total
pipe length for each chromosome. Though ne structure is essentially hidden,
the gure suggests that the landscape is deeply infected with local optima. For
example, one can imagine a hillclimber, perhaps being trapped in the far left
region of the landscape (mainly 0's) unable to escape to the global optimum far
towards the right (rather Hamming distant from it, in a mainly 1's area).
:::

:::

Fig. 8. Solution space for the 20-connection problem


Search space size for the 30 and 40 connection problems was more realistic
with respect to real problem cases, and also (of course) not feasibly amenable to
exhaustive search.

Experimental Setup Unless otherwise stated, each individual trial of an algorithm ran for a maximum of 5,000 evaluations. This is a rather modest gure, but
we felt it important to seek good performance in such a relatively small space of
time since pipe-route design optimisation problems can necessitate rather complex evaluation functions, as noted in section 2.
Standard SH was employed (random starting point; continually seek new
mutants, replace current with new only if new is better or equal to current),
using simple single-gene mutation as the neighbourhood operator. That is, a
mutation step consisted of choosing a gene at random, and ipping its value.
SA used the same neighbourhood operator, and a simple geometric cooling
schedule with starting temperature, number of iterations per temperature step,
and nal temperature set respectively at 30 the number of connections in the
problem, 50, and 0.001. These parameters arose as a fairly robust set following
much time consuming preliminary experimentation with di erent SA setups.
The GA used steady-sate reproduction, rank-based selection (as in [9], uniform crossover, a population size of 50, and the same mutation operator as SH
and SA.
Results Table 1 summarises results for trials on the 20, 30, and 40 connections

test problems. 10 trials each were made with each method, and the table records,
in column order, the best (minimum) result over the 10 trials, the worst result,
the mean result over the trials, and the number of trials (out of 10) in which an
optimal (known, in the 20-connections case) pipe-route design was found. In the

30 and 40 connections cases, number of optimal trials refers to the best-known


result for the problem.

SH
SA
GA
SH
SA
GA
SH
SA
GA

The 20-Connections test problem


Best Result Worst Result Average Result Number of Optimal Trials
202.0
218.6
207.4
0
194.4
195.7
195.2
0
192.5
194.4
192.6
8
The 30-Connections test problem
308.6
308.6
308.6
10
316.2
320.5
317.4
0
308.6
308.6
308.6
10
The 40-Connections test problem
392.9
402.3
396.2
0
397.9
406.8
404.9
0
354.3
354.3
354.3
10

Table 1. Summary of Results on test problems

On the 20-connections problem, SH performance was rather poor in relation


to both SA and GA, with SH and GA competing relatively closely, although the
GA seems clearly superior.
In the 30-connections case, SA seemed to have trouble, losing out to SH and
GA, both of which found the same best-known result in every trial. We can only
conjecture that this example may have sported a tness landscape with few or
no hard-to-escape local optima, enabling SH and GA to be steered quickly to a
result, leaving SA to unwisely, in this case, explore less promising avenues.
In the 40-connections case, the GA is clearly far superior to the other methods, with SH second-best, and SA a close third.
In each case, T-tests report the GA superior to SA with 95% con dence. In
the 20-connections and 40-connections case, T-tests also nd GA superior to SH
with 95% con dence.

4 Discussion
This initial foray into the use of GAs on pipe-route design problems appears
to show great promise. Commercial security considerations unfortunately made
it prohibitively hard to compare GA results with human expert results on real
problem examples as yet. However, we can at least report here the fact that a GA
was able to fairly reliably nd the global optimum solution in a case where the
global optimum was known (the 20-connections case), and that the GA appears
signi cantly superior to either SA or SH on this problem in larger test cases.

Extensions of this work therefore seems warranted, in particular extending


the technique to handle three-dimensional pipe-route design. In 3D, for example,
there are typically 6 Steiner points giving rise to 6 basic routes, with four main
potential escape routes around any obstacle. This leads to an obvious extension
of our representation towards dealing with the 3D case. Choosing pipe-route
diameters in conjunction with other aspects of the problem is a further necessary
extension towards real-world examples. Evidently, there is also much scope to
design problem speci c operators which should further improve results.
In summary, this study has suggested that industrial plant pipe-route optimisation can be feasibly done using genetic algorithms. We demonstrate a natural and e ective representation for the problem, and also demonstrate results
that suggest genetic algorithms are superior to either hill-climbing or simulated
annealing on these problems (at least with the representation and standard operators used here). Results point towards the potential success of extending
the technique to deal with realindustrial plant pipe-route design problems, thus
strongly supporting the work of human experts in this area.

References
1. Kang Soo Kim. Personal communication with Kang Soo Kim, Executive director
in the DAEWOO Heavy Industry Co., Seoul, Korea, August 1994.
2. Jens Lienig and Holder Brandt. An evolutionary algorithm for the routing of multichip modules. In Manner Davidor, Schwefel, editor, Parallel Problem Solving from
Nature { PPSN III, pages 588{597. Springer-Verlag, 1994.
3. Dragan A. Savic and Godfrey A. Walters. Genetic operations and constraint handling for pipe network optimization. In Proceedings of the AISB workshop. AISB,
1994.
4. J. Hesser, R. Maenner, and O. Stucky. Optimization of Steiner trees using Genetic
Algorithms. In D. Scha er, editor, Proceedings of the Third International Conference on Genetic Algorithms, pages 231{236. International Conference on Genetic
Algorithms, 1989.
5. Josep L. Ganley and James P. Cohoon. Optimal Rectilinear Steiner Tree Routing
in the Presence of Obstacles. IEEE Transactions of Computer-Aided Design of
Integrated Circuits and Systems, 1993.
6. Frank M. White. Fluid Mechanics. McGraw-Hill Book Company, second edition,
1986.
7. David E. Goldberg. Genetic Algorithms: in Search, Optimization & Machine Learning. Addison Wesley, 1989.
8. P. J. M. van Laarhoven and E. H. L. Aarts. Simulated Annealing: Theory and Applications. D. Reidel Publishing Company, 1987.
9. Darrell Whitley. The genitor algorithm and selection pressure: Why rank-based
allocation of reproductive trial is best. In D. Scha er, editor, Proceedings of the
Third International Conference on Genetic Algorithms. International Conference
on Genetic Algorithms, 1989.
This article was processed using the LATEX macro package with LLNCS style

You might also like