expert systems
The FAST Expert System
Figure 1 illustrates the flow chart for the FAST (Functional Analysis of Stone Tools) expert system computer program.
The data is then transferred into the FAST program. Each attribute, that
is the value of each variable (e.g. edge angle of 50 degrees), is used to
give an indication of motion, or hardness of material, or both, according
to a set of rules. For example the variable edge angle is divided into ranges
so that the value of the edge angle for a particular tool will fall within
one of those ranges giving the corresponding indication i.e. if the edge
angle = 42 degrees this indicates 'cutting or scraping a medium material'
(as in the example of tool 33 in figure 1). Note that the absence of a value
for a variable can be diagnostic. The data in Figure 1b shows the absence
of rounding for tool 33, together with an edge angle of 42 degrees, this
indicates a 'soft to medium material', because if the worked material had
been hard then some rounding would have been expected on a 42 degree angled
edge. Conversely the value of a variable may be non-diagnostic. For example
there are micro flakes on the ventral surface of tool 33, but as flake fractures
can occur with almost any motion and with any worked material, the presence
of these flakes is non-diagnostic. With retouched edges the value 'retouch'
is entered because of the difficulties of separating use wear fractures
from retouch.
This process is repeated for each attribute. The program automatically assesses
the attributes and enters the relevant indications into two cards, one of
which contains information concerning macro observations ( use wear seen
by eye and with low magnification, fig. 6)
The syntax for these variable rules is very simply and takes the form
of,
IF [condition] THEN PUT [indication].
For example; IF (edge angle <30 degrees) THEN PUT (cutting soft material).
The rules may be more complex involving 2 or more conditions to take account
of the interaction between different variables.
For example; IF (fractures are absent) AND (edge angle >30 and <60)
THEN PUT (medium material).
The parameters contained in these rules are derived from observations of
experimental tools. The indications are then counted, again according to
a set of rules.
For example;
EACH VARIABLE COUNTS AS TWO POINTS [except thickness which has a maximum
of 1]. This is because thickness only has two values <4 mm or > 4
mm and is not very discriminatory and consequently less important. Therefore
it carries less 'weight'.
IF EITHER SURFACE HAS TWO INDICATIONS THEN EACH COUNTS 0.5 POINTS UNLESS
OTHER SURFACE IS "retouch" "no polish" OR "no effect"
THEN THEY COUNT ONE POINT EACH. If an indication contains two alternatives
such as 'SOFT/MEDIUM for micro rounding (as in Figure 1g) then SOFT would
receive 0.5 points but doubled to 1 point because the other surface is retouched.
"NON DIAGNOSTIC" COUNTS NO POINTS, is self explanatory.
"GROOVING" , "WHITTLING" and "PERCUSSIVE"
scores are doubled. This is because the attributes that indicate these motions
are more diagnostic than others and so this a method of weighting the variables.
The results of the counting rules are entered as SCORES into the interpretation
card. In the example, this gives 12 indications of scraping, 1 of cutting,
5 of a soft material, 8 of a medium material and 1 of a hard material (fig.
1h). Then the function rules are applied. For example;
IF "cutting" <4 AND "scraping" >8 AND"grooving"
<2 AND "whittling" <2
THEN PUT "SCRAPING"
IF "soft" >4 and <8 AND "medium" >0 and <2
AND "hard" = 0
THEN PUT "SOFT"
IF "soft" <6 AND "medium" >5 AND "hard"
<4
THEN PUT "WOOD"
More complex rules involve combining motions with materials, and in certain
cases also including morphological information concerning the tools.
IF "soft" >2 and <6 AND "medium" <8 AND "hard"
<2 AND MOTION "whittling" OR "boring/drilling" OR
"grooving" OR "chopping/adzing"
THEN PUT "HIDE"
This rule is constructed in this way because whittling, boring, drilling,
grooving, chopping and adzing are motions unlikely to be used on hide.
IF "soft" =0 AND "medium" >3 AND "hard"
>8 AND MOTION "whittling" OR "cutting" OR "piercing"
OR "chopping/adzing" OR "grooving" AND SUBTYPE "facet"
(when referring to a burin)
THEN PUT "STONE"
This rule is constructed in this way because whittling, cutting, chopping
and adzing are unlikely motions to be used on stone and grooving stone is
more likely to be carried out with the burin 'bit' rather than the 'facet"
If the scores for motions and materials fall within the parameters in the
program then an interpretation will be made of motion, hardness of material
and precise worked material. In the case of the example tool 33, the program
gives SCRAPING a SOFT/MEDIUM material probably HIDE (fig. 8), which is correct,
as tool 33 was an experimental tool used in a blind test (Grace
et al., 1988).
If the scores do not fall within the parameters for motion, hardness
or worked material then the program gives 'INSUFFICIENT DATA'. This will
apply if there is insufficient use wear on the tool to be diagnostic or
if the use wear is not consistent with a particular use. That is, it does
not match the use wear of tools in the reference collection of experimental
tools from which the parameters were derived. This means the program can
suggest a material that has not been studied by experimentation and so is
not included in the program. Tool 44 used on cortex being an example (see
Grace 1989).
Prior to the development of the expert system computer program the interpretation
of each attribute had to be done by assessing the information and the complex
interrelationships between attributes in one's head, as it where. FAST carries
out this process automatically. This not only speeds up the process but
makes it completely consistent as the same set of rules are applied each
time. The 20 tools used in the last blind test carried out at the Institute
(Grace et al., 1988) were used to determine
the parameters by which the rules were applied in order to make the functional
interpretations. That is, the data that was recorded for that blind test
was used as the training data for developing the program. The efficacy of
the program is demonstrated by it achieving a result of 18 out of 20 correct
interpretations of precise worked materials. The two tools that were not
correctly identified were tool 38 for which the computer gave "insufficient
data" (used on bark) and tool 44 which was designated "insufficient
data" (this was used on cortex which was not programmed into the computer).
The same scoring system as used in the blind test was applied, therefore
to achieve a point the precise worked material had to be identified. If
the tool was used on antler than only an answer of antler was awarded a
point, not alternatives like bone/antler. This 90% success rate is a significant
increase on the result achieved by any of the analysts in the original blind
test, the maximum score achieved being 60%. However, as the blind test data
was used to develop the program this high rate of success is misleading.
The real test of the program is when a completely new set of data is used.
The first 10 tools used in blind tests at the Institute (Newcomer
et al., 1986) were observed and the data recorded and then used to test
the program. Of the 10 test tools, one was unused and another was used as
a projectile point that struck unknown material, leaving eight precisely
known materials that the tools were used on. Of these 8 tools, the precise
worked material on which 6 of them had been used was identified by the program.
The two not identified were tool 2 for which the computer gave "insufficient
data" (actually used on shell), and tool 10 for which the computer
gave "antler" but which was used on wood. In the original test
only two of these eight were correctly identified. Though the function of
these tools was known, this was a blind test as the computer did not have
this information. Every time the FAST program is run constitutes a blind
test.
Since the development of FAST another expert system for the analysis of
use-wear on stone tools has been developed. This program is called Waves
(Wear Analysing and Visualising Expert System) see Dries
1994. The program is similar in structure to FAST as data are entered
by clicking on the appropriate observation from a multiple choice menu for
each variable, and IF-THEN rules are applied in order to match combinations
of observations against experimental data. The interpretations are presented
in the form of probabilities and a certain level of data is required in
order to make an interpretation