Computational Intelligence - August 2014 - 80

CA (20th Generation)-Residential Clusters of Clusters

350
300
250

R9.1
R9.2
R9.3

200
150

R8.2
R8.3 R8.1
R8.4
R5.4
R5.1
R8.5
R5.2
R5.3
R6.2 R6.4
R6.5
R6.3
R7.4 R7.2 R7.5
R4.3 R4.2
R6.1
R7.1
R7.6
R4.1
R2.3
R7.3
R2.1
R2.4
R3.2
R3.1
R2.5
R2.2

100

R1.3

R1.4
R1.1

0
180

200

220

240

260

280

300

R1.2

320

Road System

Minor Gate

Reservoir

Major Gate

Main Plaza

340

360

Figure 17 The three non-elite residential regions (clusters of clusters) produced by the best
Cultural Algorithm run. The total fitness value of this part of the overall plan is 87%.

knowledge source and can influence the
population in the next generation. The
minimum and maximum number of
proposed regions (cluster of clusters) of
the best plans are used to update the
normative knowledge. In addition, the
center locations of the regions in the
best plans are used to update the topographical knowledge.
In the following sections the plans
produced from the best runs of the Cultural Algorithm are described for each of
the three functional classes (non-elite residential, crafts, and elite-residential). For
each class, the learning curve for that best
run is given, along with the contents of
the cluster produced, and the physical
distribution of those cluster of clusters
over the site.
A. Best Run for the Non-Elite
Residential Class

IEEE ComputatIonal IntEllIgEnCE magazInE | august 2014

resiDenTial

eliTe

mano &
meTaTe

liThiC

CeramiC

shell

obsiDian

norTh sTD

Fitness Value (Average Regional Scores)

easT sTD

CounT

norTh

easT

region

The Cultural Algorithm was run 50
times, and the learning curve for the best
Table 18 The regions contained in the best non-elite residential plan are given in the
run is given in Figure 16. The fitness
table. The coordinates (East and North) for its central location and their standard
value for the best plan plateaus at around
deviations are provided along with the percentage of the total for the period for each
20th generation which means that the
category of terrace activity. The regions were extracted after 20 generations.
resultant plan achieved an 87% score on
the set of weighted constraints. The Cultural Algorithm produces an s-shaped
learning curve for the best plan until it
reaches its maximum at 87% as shown in
Figure 16. The average score for all exe1
310
105
122
116
174
0.66 0.76
0.43 0.53
0.06 0.49 0.35
cuted plans at each generation improves
2
231
212
132
143
163
0.18
0.16
0.26
0.13
0.47
0.26
0.39
slightly in a somewhat a linear fashion
3
303
240
104
159
238
0.07
0.00 0.2
0.23
0.47
0.11
0.31
and reaches a score of around
65% when execution is termithe domain knowledge for that
1
nated at 60th generation. The
functional category.
0.9
minimum value also exhibits a
In every generation of a run,
slight increase by the end of 60th
each agent searches for a city
0.8
generation. A portion of this
plan. Each plan is composed of a
0.7
increase is due to the improvenumber of regions. Each regions
0.6
ment in the overall minimum
has several attributes that reflect
0.5
that is produced as more inforits content, context, and structure
mation about how to satisfy plan
in terms of the building blocks.
0.4
constraints is incorporated into
At the end of each generation,
0.3
the belief space during the run.
each agent's plan is evaluated
Max (Best)
0.2
Ave
The minimum is close to 55% at
using the fitness function deMin
the end of the run.
scribed in Section IV in terms of
0.1
The original fitness value was
the various weighted rule sets.
0
0
10
20
30
40
50
60
61% in 1st generation and
The best 20% of the population
Generation
increased to 87% by 20th generaplans are chosen based on their
tion. The relatively high initial
fitness scores, and used to update
Figure 18 The learning curve results for the best craft run out of
score reflects the quality of
the knowledge space. The best 50 runs where each run takes 60 generations. The fitness value for
knowledge initially presented to
agents are stored in the situational this run is 93%.

Table of Contents for the Digital Edition of Computational Intelligence - August 2014