letric 
1 the 
/ and 
jeing 
najor 
step, 
and 
strate 
bility 
con- 
xper- 
gn of 
ch as 
ould 
‘Wal- 
NS is 
with- 
1 and 
'con- 
k de- 
s in 
deci- 
, Le. 
con- 
in are 
alisa- 
[hese 
ropri- 
ed in 
ftwo 
ames, 
ge in 
diag- 
efine- 
pe as 
adens 
and the understanding of the task improves. The consid- 
erations in this paper should be viewed in this light. 
2 ON BUILDING AN EXPERT SYSTEM 
Building an ES is a complex, ill-structured, and inherent- 
ly experimental activity: "there is little chance every- 
thing can be figured out beforehand" (Buchanan et al, 
1983). Nevertheless, as a guide, this process can be di- 
vided into 5 steps, namely problem identification, con- 
ceptualization, formalization, implementation and 
testing, as shown in Figure 1. The identification step en- 
tails selecting a suitable task for ES development, defin- 
ing the related problem domain, establishing project 
goals, and characterizing the important stages of the task. 
In defining the application domain for CONSENS to be 
in support of a measurement robot, this step has already 
been addressed. In conceptualization, the key attributes 
of the task and domain are made explicit. To this end, the 
knowledge employed by experts in reaching solutions in 
the problem domain may be transcribed into flow charts, 
diagrams, lists etc., which serve to expose the strategies, 
relations and information flow. Formalization of knowl- 
edge constitutes a mapping of the key concepts, sub- 
problems and information-flow characteristics isolated 
during the conceptualisation stage into more formal, 
knowledge-engineering representations. The output of 
this step is a partial specification for building the ES. Im- 
    
   
  
  
    
  
Latjeremenes reformulations 
ap leformulations ............. 
ES re-designs mencssssrionrees 
ailes esesesess refinements ............... 
  
  
  
Figure 1 Stages in the building of an expert system (af- 
ter Buchanan et al, 1983). 
plementation involves mapping the formalized knowl- 
edge into the representations supported by the selected 
ES development tool. The last step, testing, involves 
evaluating the performance of the ES, e.g. against some 
case studies for which solutions exist. With this test step, 
feedback loops (dashed lines in Figure 1) in the form of 
refinement of the ES, redesign of the knowledge repre- 
sentations, or reformulation of the task conceptualiza- 
tion, indicate iterative revision of the ES (Buchanan et al, 
1983; Dym, 1987). 
3 ON CONCEPTUALISING CLOSE-RANGE 
NETWORK DESIGN 
In coarse terms, close-range network design is the proc- 
ess by which the goal of precise, reliable and economic 
object measurement is achieved through configuration of 
a suitable photogrammetric triangulation network. As 
shown in Figure 2, this process can be conceptualised in 
(at least) three different ways: (i) using the network de- 
sign classification scheme introduced by Grafarend 
(1974); (ii) in terms of the design-by-simulation strategy 
employed by design experts; and (iii) in terms of generic 
problem-solving processes. Each conceptualisation as- 
sists in understanding the nature of the task. 
    
      
   
  
     
    
    
     
Measurement Criteria: 
* precision 
» reliability 
* economy 
    
    
   
   
  
   
   
    
    
  
    
    
     
  
  
task-based: design-by- generic 
«ZOD simulation Bronte 
. strategy solving 
FOD processes 
«SOD 
Goal: 
network design 
Figure 2 Network design can be conceptualised in terms 
of tasks, solution strategy or generic problem 
solving processes. 
3.1 Grafarend’s Classification Scheme 
According to Grafarend’s (1974) classification scheme, 
general network design requires solving four major tasks, 
commonly known as zero- (ZOD), first- (FOD), second- 
(SOD), and third-order (TOD) design. In relation to 
close-range photogrammetric applications, these tasks 
can be defined as: 
ZOD: defining a datum for the measured object 
points; 
FOD: configuring an optimal imaging geometry; 
SOD: adopting a suitable measurement precision for 
the image coordinates; and 
TOD: network densification, although largely irrele- 
vant to the vast majority of close-range net- 
works (Shortis and Hall, 1989). 
Many considerations pertaining to ZOD, FOD and SOD 
for photogrammetric networks can be found in the litera- 
ture (e.g. Hottier, 1976; Fraser, 1984; Grün, 1985; Shor- 
tis and Hall, 1989; Fraser, 1992). Because of the 
dependencies between each of these tasks (e.g. ZOD in- 
volves the choice of an optimal datum given the network 
design and the precision of the observations (Fraser, 
447