A.V.Gavrilov, Novosibirsk
State Technical University,
630092 Novosibirsk 92,
K.Marx av., 20,
email:
avg@silver.nstu.nsk.su
The actual problem of AI
development is combination of traditional logic proceeses in expert systems
(ES) and associative adaptive processes in artificial neural networks (ANN).
Two kinds of such combination are described in this paper.
First of ones is based on
constructing of expert system from two parts : the logic expert system ( LES )
and the program model of ANN, as in brain. This kind of this “two-semi-sphere”
architecture is implemented in the expert shell [1].
Second one is based on the
implementation of any properties of neurons in usual method of knowledge
representation, such as methods of rules, frames, semantic nets. In particular, the rules are choosed for the
implementation in expert system for professional orientation - choosing of
profession and diagnostic of abilities for defined profession.
Architecture of the expert shell based on the artificial
neural network.
LES deals with logic processing and user
interface.
This subsystem may influence on ANN in consider of results
of analize of it outputs.Artificial Neural Network are de-signed for set of associative links
between facts or hypo-thesis.
To choose of architecture
principles of LES we must have
in mind that the different complex methods of
organizing Expert Systems ( such as the representation of fuzzy knowledge ) are
the consequance of absence of such device as Artificial Neural Network. So it seems we may utily the simplest principles
in the architecture of LES. The
main demand to this architecture is the orientation to parallel asynchronouse deduction with
possibility to remove obtained conclusions.
The following architecture
principles of LES may be formulated:
·
the mechanizm of black board for
connection between sources and
receivers of knowledges;
·
the knowledges transmitted
across the black board are facts
and hypothesis formed as triplet “object-attribute-value” or the part of this
triplet ( for example, “object” or “atribute-value”);
·
the sources and
receivers of facts and hypothesis
are rules with representation of indetermination;
·
one of sources and receivers
of facts is
Artificial Neural Network;
·
to send of punishment signal to ANN for actuaating of adaptation (
changing of state) of it the special fact is used;
·
the punishment signal
may be the result of succesful applying of any rule or result of analyzing
of facts generated by ANN;
·
to transfer facts to inputs of
ANN the dictionary
and coding/decoding are used.
Any rule of LES is started
by Manager of the Black Board ( MBB ) after appearance ( from
another rule or ANN ) of corresponding fact or
hypothesis.
In this Expert
System the architecture of Artificial Neural Network proposed in [2,3]
is used.
This architecture differ from most known ones by view on neural
element as the simple perceptron learning to recognize the
simple image - binary
input code, but
not simple threshold element.The Neural Element (NE) is
described by threshold, key, input links with another NE and input link with
another NE, from which one gets the punishment signal.
The neural network works in descrete time, i.e.
for time t the outputs of all elements are determined by inputs in t-1.
Two software platforms are
used for an implementation of this Expert Shell: PDC-Prolog in MS DOS and
Borland C++ in MS Windows.
Architecture of the expert system for professional
orientation.
This expert system solves following
tasks:
·
the testing of existing psichological haracterictics,
·
the choosing of more suitable profession,
·
the testing of ability to deal with any profession,
·
the explanation of reasons for
refusal from any profession,
·
the choosing of any analog profession.
To include the parts of any
psichological tests in know-ledge base the nontradional kind of intepretation
of the rules was applied in this ES. This form of the interpretation is based
on the threshold and the addition with weghts for making of conclusion. The
rule in this ES is described by the following syntacsis:
<Condition i> <CNF>
<Conclusion j> <CNF>
Here CNF in condition has
semantic depending on the kind of interpretation of rule. In case of usual
logical interpretation CNF is the threshold of the confidence factor and
condition is true if the confidence factor of corresponding fact greater then
it.
In case of addition
interpretation CNF is the weght of the condition and the conclusion is formed
always with CNF calcu-lated from confidencies of conditions. The combination of rules with 2-nd and 1-st
kinds of inter-pretation is modelling the neuron with distributed functions
between some rules.
1) Gavrilov A.V., Novickaja J,V. The Expert Shell
based on the Artificial Neural Network. - NITS’94, Penza, 1994, pp. 43-45.
2) Gavrilov A.V. An Architecture of Neurocomputer
for Image Recognition& - Latvian Signal Processing International Conference
(LSPIC’90).Proceedings.- Riga,1990,pp. 306-308.
3) Gavrilov A.V. An Architecture of Neurocomputer
for Image Recognition. - Neural Network World, 1991,N.1, pp.59-60.