°p lev- 
s such 
play a 
Iristics 
espect 
fied in 
ork 
ceptu- 
ses. In 
rategy 
ithmic 
mance 
cation 
ribing 
twork 
  
A 
EL 
1 and 
S) 
  
of ge- 
ns of 
)blem 
hesis, 
n, to 
986). 
servations. Typically they relate observed behav- 
ioural irregularities with underlying causes, using 
one of two possible techniques. The first method 
essentially employs a table of associations be- 
tween behaviours and faults (generally heuristic 
knowledge) The second method combines 
knowledge of system design with knowledge of 
potential flaws in design, implementation, or 
components to generate candidate malfunction 
consistent with the observations (model-based 
reasoning) (Hayes-Roth et al, 1983). 
* The prescribing process, in the context of network 
design, is a form of design: a previously designed 
configuration is corrected to overcome diagnosed 
faults. 
This conceptualisation is useful insofar as it provides a 
means by which the experience gained in the building of 
other expert systems can be applied to the current prob- 
lem. To this end, reasoning strategies for each step in the 
design of networks by simulation can be inferred from 
the strategies employed for the related generic problem- 
solving processes. With reference to Table 1: 
* The large solution space common to complex de- 
sign problems is often reduced by experts by (of- 
ten heuristically) breaking it down into sub-goals, 
these being related to the attributes of the design 
in it’s final, desired state. This acts to reduce the 
search space to a manageable size. (The design- 
by-simulation strategy presented in Figure 3 is an 
example of this.) Consequently, in rule-based 
ESs, search can be limited by using a goal-direct- 
ed, backward-chaining reasoning strategy (Dym, 
1985; Oxman and Gero, 1987). 
* The structure of the search space for diagnostic 
problems is most often the reverse of that for de- 
sign. The goals - identified faults - are unknown 
and must be inferred from the observational data 
available e.g. from an evaluation of a network de- 
sign, in this case. To this end, a data-driven, for- 
ward chaining reasoning strategy is most 
appropriate (Oxman and Gero, 1987). 
  
  
  
  
  
  
Network design Generic problem- Reasoning 
task solving processes strategy 
initial design design BWD chaining 
performance algorithm procedural 
measures 
design diagnosis diagnosis FWD chaining 
FOD, SOD prescribing BWD chaining 
corrections 
  
  
  
  
  
Table 1: Reasoning strategies for network design 
e The prescribing of corrections to a design em- 
ploys the same reasoning strategy as for design. 
To this end, design goals are set with respect to 
the diagnosed faults. 
The use of forward chaining in the diagnosis of network 
designs is exemplified in Section 4.2 below. 
3.4 Example: Conceptualizing Diagnosis in Network 
Design 
The decision-making processes within each of the indi- 
vidual steps (initial design, etc.) of the design-by-simula- 
tion strategy are not made explicit by Figure 3. In this 
section, a small portion of network diagnostic knowl- 
edge is identified and provisionally conceptualized. This 
example serves to: (i) illustrate the role and importance 
of heuristic knowledge in network design; and (ii) to pro- 
vide the basis for investigating appropriate representa- 
tions for network design knowledge (the topic of Section 
4). 
The objective of diagnosis in network design is to identi- 
fy the faults which cause a network to fail set precision, 
reliability and economy criteria. As depicted in Figure 5, 
  
  
  
  
   
   
  
  
   
  
  
  
  
  
   
evaluation 
data Task: 
identified 
network faults 
diagnosis 
measurement 
criteria 
  
  
  
Figure 5 Input/output model of the network diagnosis 
task. 
input to this task consists of evaluation data (e.g. the var- 
iance-covariance matrix for the object point coordinates 
obtained from bundle adjustment) and the measurement 
criteria (e.g. rms precision of point determination to be 
reached). As described below, expert knowledge in the 
form of heuristics is used to identify faults from these in- 
puts. 
Heuristics can be defined as the rules-of-thumb and em- 
pirical associations that, gained through experience, ena- 
ble experts to make educated guesses when necessary to 
recognise promising approaches to problems (Waterman, 
1986). From the literature and interviewing network de- 
sign experts, a number of heuristics with respect to the 
first step in network diagnosis - "criteria satisfied" (see 
Figure 3) - can be identified: 
* As precision measures are not of much value if 
the reliability of a network is unacceptable (Grün, 
1980), each design should be tested for reliability 
before precision. 
* Assuming that the number of non-parallel rays in- 
tersecting at a point can be used as a rough meas- 
ure of point determination reliability, a first test of 
reliability is that each target point should be inter- 
sected by at least 4 non-parallel rays (Grün, 
1980). 
449