EXPERT SYSTEMS
"an expert system is a computer program which uses non-numerical
domain-specific knowledge to solve problems with a competence
comparable with that of human experts" ((Doran
1988).
EXAMPLE of a simple expert system to
identify 20Kroner, 5Kroner and 1Kroner Norwegian coins.
The first step is to identify the variables-
SIZE
COLOUR
DECORATION
Then assign a range of values for each variable-
SIZE
diameter is >25mm
diameter is <25mm
COLOUR
silver
bronze
DECORATION
head
crown
ship
lion
Then rules are constructed that identify the coins by combinations
of attributes (attributes are the particular value of a variable
e.g., a coin can have the attribute of being silver in colour,
or decorated with a lion).
RULES:
IF SIZE > 25 and COLOUR is bronze and DECORATION is ship THEN
coin is 20K
IF SIZE < 25 and COLOUR is silver and DECORATION is crown THEN
coin is 1K
IF SIZE >25 and COLOUR is silver and DECORATION is lion THEN
coin is 5K
These rules can be simplified to extract the distinguishing features
of the coins (or stone tool types in LITHAN or tool function in
FAST)
IF SIZE > 25 and COLOUR is bronze THEN coin is 20K
IF SIZE < 25 and COLOUR is silver THEN coin is 1K
IF SIZE >25 and COLOUR is silver THEN coin is 5K
When dealing with incomplete specimens which have missing attributes,
'fuzzy logic' is employed in order to make probability statements.
(see
a tutorial on fuzzy logic)
IF SIZE > 25 THEN add 1 into coinA
IF SIZE < 25 THEN add 1 into coinB
IF COLOUR is bronze THEN add 1 into coinA
IF COLOUR is silver THEN add 1 into coinB
IF DECORATION is ship THEN add 1 into coinA
IF DECORATION is crown THEN add 1 into coinB
From these 'scores' fuzzy logic probabilities can be assigned-
IF coinA = 3 Then coin is a 20Kroner (fuzzy logic probability
of 1)
IF coinB = 3 Then coin is a 1Kroner (fuzzy logic probability of
1)
IF coinA = 2 Then coin is PROBABLY a 20Kroner (fuzzy logic probability
of 0.7)
IF coinB = 2 Then coin is PROBABLY a 1Kroner (fuzzy logic probability
of 0.7)
Fuzzy logic probabilities are not employed in LITHAN because tool
types are based on mutually exclusive categories, but are used
in FAST because use wear features often overlap and are therefore
not mutually exclusive categories. see Use-wear.
The first major advantage of using an expert system for lithic
analysis is the act of writing it. "The process of developing
an expert system has an indirect benefit also since the knowledge
of human experts must be put into an explicit form for entering
in the computer. Because the knowledge is then explicitly known
instead of being implicit in the expert's mind, it can be examined
for correctness, consistency and completeness. The knowledge may
then have to be adjusted or re-examined which improves the quality
of the knowledge." (Giarratano
and Riley 1989, 5).
The expert system approach is essentially looking for patterns
in complex dynamic phenomena that have proved to be beyond standard
quantification techniques. For example 'ship decoration' cannot
be quantified. The outcome of the expert system is a probability
statement concerning the tool type or function that is most consistent
with the observations. The interpretations are made according
to the balance of indications given by the expert system rules
and based on the observation of all features.
Expert systems are not intended to replace human experts. For
example, the recognition of retouch on stone tools as opposed
to edge damage (from spontaneous retouch, trampling, post depositional
movement, etc.), is dependent on the analyst's experience and
in particular on experimentation, involving not only observation
of experimental and archaeological tools , but also an appreciation
of the mechanics of making and using stone tools.
Expert systems ensure that interpretations are consistent and
comply with the tenets of scientific method. For example, one
definition of scientific schemes describes such expert systems,
"... scientific schemes are explicit i.e., the rules and
the way they are to be applied are spelled out with sufficient
clarity and in enough detail that they can be used by anyone.
... such a set of rules can be encoded in a computer program...
" (Casti 1993,29).
The expertise gained over many years of research is made available
to less experienced practitioners. One of the features of expert
systems is that, "The expert system may act as an intelligent
tutor by letting the student run sample programs and explaining
the system's reasoning." (Giarratano
and Riley 1989,5). As an expert system models the behavior
of an expert (hence the name), the incorporation of such expert
systems into teaching programs enables students to understand
the reasoning processes of the expert rather than simply learning
the outcome of the reasoning. As the rules that operate the expert
system are derived from a number of sources the expertise of many
researchers is incorporated into the program. The LITHAN and FAST
programs are currently being used as part of a teaching program
for lithic analysis.
The use of expert systems has a number of advantages over other
techniques.
Increased consistency and standardisation. The development of
an expert system means that the observational techniques have
to be systematised and the rules provide a base from which results
can be assessed.
Different analysts using the same program will obtain the same
results. This has been repeatedly confirmed during instruction
in use-wear analysis when several students have independently
analysed the same experimental tools and all interpreted the correct
function of the tool using FAST. Often students enter different
observations, due to inexperience, but the flexibility of the
program (in particular the 'fuzzy logic' aspects) allows for this
so that some variations in observations can be accommodated.
Analysts working on different material can use the same program.
As demonstrated in a recent study of lithic material from Tehuacán
and Oaxaca in Mexico which involved using local material in replication
experiments (Hardy 1993)
The rules and procedures for expert systems can be continually
being updated in order to improve and refine the analytical procedures.
For example, since the FAST program has been in use in Norway
(Ballin & Jensen 1995)
a large number of experiments have been carried out on fish. The
information gained from these experiments has been incorporated
into the rules of the FAST expert system making the identification
of fish processing more accurate. Subsequently these new rules
have helped in identifying fish processing during current research
on Neanderthal associated material from Amud cave, Israel.
The use of rule based expert systems is a practical approach to
lithic studies that bridges the gap between processual and post
processual archaeology. The key here is rules; not laws which
are inviolate, but rules that can be changed and indeed are always
changing in a reflexive relationship allowing the expert system
to accommodate new information.
The rules of the expert system are subjective, but they are explicit
in that they are written down and incorporated into the computer
program. The observations are defined and the rules are explicit
therefore anyone can produce the same results, so that though
the system is subjective it is consistent when different subjectivities
(i.e. different individuals) use it. The acceptance of the assumptions
on which the program is based leads to consistency, and direct
comparability between results produced by different people; this
fulfills the basic requirements of objective data within the consensus
reality of mutual users of the program. Therefore expert systems
can extract objective-like data, but the complexity of the dynamic
process is retained and the data is produced in the form of probabilities
that can be compared as if they are objective data within a defined
consensus reality.
Expert systems are so called because they are designed to model
the behaviour of a human expert. So they are modeling human behaviour,
in fact an individuals behaviour. By extension expert systems
can be used to model the more complex behaviour of societies.
A series of programs that input the results of each individual
program into another program further up the hierarchy is being
developed. Not only must the interpretations be consistent with
use-wear analysis and lithic programs, but non-lithic material
such as the faunal assemblage, environmental evidence and spatial
information from the site and any chronological evidence.
Alternative interpretations can be modeled with expert systems
so rather than postulating a theory and then testing it, a number
of alternatives can be tested and matched against the data simultaneously.
"an expert system is a computer program which uses non-numerical
domain-specific knowledge to solve problems with a competence
comparable with that of human experts" (Doran
1988).
"The process of developing an expert system has an indirect
benefit also since the knowledge of human experts must be put
into an explicit form for entering in the computer. Because the
knowledge is then explicitly known instead of being implicit in
the expert's mind, it can be examined for correctness, consistency
and completeness. The knowledge may then have to be adjusted or
re-examined which improves the quality of the knowledge."
(Giarratano and Riley 1989,
5).
Rules are then applied to interpret the blank type, knapping
technology, hammermode, amount of cortex and the 'type' of tool.
Blanks can be blade, bladelet, flake, chip, fragment or chunk.
Knapping technology can be blade, flake or Levallois. Hammermode
will be soft or hard, and cortex is broken down into 4 categories
dependent on the percentage of surface that is cortical, (this
information is useful in the reconstruction of reduction strategies).
The expert system will then display it's findings on the interpretation
card.
In the case of tool 33 this gives a non-cortical morphological
flake that was made using
a blade technology with soft
hammer and is an end
scraper
Often there is insufficient data to identify such categories as
knapping technology or hammer mode, particularly when the tools
are broken and the proximal end is missing. In such cases they
will be designated 'indeterminate'
Examples of rules:
BLANK TYPE: if length/width ratio >2 and width <12 mm. then
put "BLADE LET"
TECH TYPE: if platform Thickness <5 and ButtType = "prepared"
and Sides = "parallel" and Ridges = "parallel"
then put "TECHBLADE"
HAMMERMODE: if percussionCone = "no cone" and butt =
"un-lipped" and bulb = "diffuse" then put
"SOFT HAMMER"
TYPE: if diff (length - width) > 0 and distalRetouch = "DISTAL"
then put "END SCRAPER"
General categories like endscraper are further subdivided by applying
secondary rules.
1) if endForm = "ROUND" then put "END SCRAPER"
2) if endForm = "CARINATED" then put "CARINATED
END SCRAPER"
The actual rules run to some 30 pages of programming in order
to cover as many alternatives as possible. These rules are being
constantly updated and expanded. The main advantage of the LITHAN
program is consistency, in that anyone using the program will
obtain the same results, eliminating some of the idiosyncrasies
that often occur with individual typologists. Also years of experience
of a number of typologists are encapsulated in the program so
that this accumulated experience is made available to the novice.
There are sub-routines for special categories:
Examples:
CORES
BURINS
ARROWHEADS
MICROLITHS