IF the hypothesis all_points_sufficient_rays is TRUE & 
IOPTI.Max convergence, angle » "limit" 
THEN all points strong geometry 
Rule 5: 
IF the hypothesis all points sufficient rays is TRUE & 
IOPTI.Max, convergence angle «- "limit" 
THEN some points poor geometry 
AND Add these IOPTI to class IWEAK PTI 
IF the hypothesis all points strong. geometry is TRUE 
IOPTI».Precision «- Precision Criteria 
THEN network passes precise, criteria 
Rule 7: 
IF the hypothesis all points strong. geometry is TRUE 
& 
IOPTI».Precision > Precision Criteria 
THEN network fails precision criteria 
The syntax of these rules may appear somewhat cryptic, 
but can be easily understood with the help of an example. 
In Rule 5, the hypothesis some points poor: geometry. 
will only be set true (confirmed) by the inference engine 
of the ES if all points sufficient rays and the conver- 
gence angles of some object points are poor. Notation of 
the form lclass| used refers to the frame-represented 
structure of the knowledge. By use of pattern matching, 
this structure can be reasoned with in the rules. For ex- 
ample, in Rule 5 IOPTI.Max convergence angle requires 
the ES to test the maximum convergence angle of each 
object point in the network. Those points failing the test 
criteria are Created as instances of a second class, 
WEAK PT, similarly by pattern matching. Additional di- 
agnostic rules need then only address the points in this 
latter class when searching for the cause of the poor con- 
vergence angles. 
As implied from Table 1, forward chaining is an appro- 
priate reasoning strategy in diagnostic tasks. In this strat- 
egy, the inference engine applies known data to the 
conditions of each rule (LHS) in order to determine the 
value of hypotheses (RHS). Thus, if the hypothesis all_- 
points sufficient rays is true, rules 4 and 5 will be evalu- 
ated because they contain this hypothesis as a condition. 
Sequences of rule applied by an ES to reach conclusions, 
such as these, are termed inference chains (or paths). Fig- 
ure 11 shows the inference paths obtained from the rules 
listed above. Expressed graphically in this manner, it can 
be clearly seen that (i) with forward chaining, the shape 
of the search space is exploited - branches of the decision 
tree containing knowledge not relevant to the current 
problem are cut off at an early stage; and (ii) the reason- 
ing of the ES corresponds to that of the human expert in 
deciding whether or not a network satisfies measurement 
criteria. 
Note finally that inference chains can be used by the ES 
to explain how it reached a particular conclusion. This 
retrospective reasoning mechanism is the most common- 
ly implementation of explanation in ESs (Waterman, 
1986). If, for example, an ES with the rules listed above 
were to be asked to explain why it concluded that a net- 
work satisfies precision criteria, the response may be: 
As the number of ray at each point is > 4 
There is evidence that all points have sufficient rays (Rule 
2) 
And as the convergence angle at each point is o.k. 
There is evidence that all points have strong geometry 
(Rule 4) 
And as the precision of each point is better than the criteria 
There is evidence that the network satisfies the precision 
(Rule 6). 
vetere tenente ent forward chaining -————7—7—7—7—-—-——7—7—4——————q- 
AP" 
    
Figure 11 The inference path associated with the diag- 
nostic rules. The appropriate reasoning strate- 
gy is forward chaining. Ri refers to Rule i. 
This information is not only useful in debugging the 
knowledge base of the ES, but can also assists non-ex- 
perts in understanding the reasoning involved in the de- 
sign of photogrammetric networks. The ES can thus be 
used as a training tool for non-experts. 
5 SUMMARY 
A brief introduction into two of the tasks - conceptualiza- 
tion and formalization - involved in building CONSENS, 
an expert systems for close-range network design was 
provided in this paper. By conceptualizing the design-by- 
simulation strategy used by experts into generic design, 
diagnosis and prescribing problem-solving processes, ap- 
propriate reasoning strategies for the various tasks of this 
strategy were established. 
Some heuristic knowledge involved in the diagnosis of 
networks was conceptualised into a decision tree. It was 
shown that this decision tree could be formalized into 
two standard knowledge-engineering representations - 
rules and frames. Structural (hierarchical) knowledge in 
network design e.g. the camera station, object point, im- 
age point and image elements forming the network itself, 
    
    
   
    
  
   
   
     
  
  
   
   
    
   
  
  
  
  
  
   
   
   
   
   
   
  
   
    
   
   
   
   
   
  
    
    
   
   
   
   
   
   
   
po qu pem 
m4 4 — (CS b o m go bed 
Pog ya EN MAMA E | e 
dud