Section 5: Concluding
remarks
In a recent paper Langley
& Simon (1995) pointed out that traditional artificial intelligence has
fielded hundreds of expert systems in industry. Five different paradigms for
machine learning were discussed. These are; (1) neural networks, (2)
instance-based, (3) genetic algorithms, (4) rule induction and (5) analytic
paradigms. Langley & Simon then go on to illustrate recent advances in
machine learning.
However, only two of the
examples where from a non-engineering domain. In the first example, an expert
system made more accurate weather forecasting than numerical models. Langley
& Simon also reported work completed by Muggleton et al (1992) where
analytic methods were used to gave predictive rules for conformation (folding
patterns) of amino acid sequences.
The work by Dubchak &
Muchnik (1995) uses some standard methods from case based reasoning as well as
associative neural networks. Their methods are very similar to those employed
by Zabezhailo et al (1995, 1996) and are in fact from the "Russian"
point of view. The work by Zabezhailo has been used to predict the properties
of pharmacological agents from a structural activity relationship (SAR)
analysis associated with molecular graphs of carcinogens. Both groups use
empirical research to classify biochemical dynamics using a theory of types.
The properties of types of events under study, protein conformation in the
first case and carcinogenicity in the second, are encoded into a database. A
second database of structural components was developed and a bi-level theory of
relationships constructed at the situational level. Dubchak & Muchnik used
a neural network to encode the structural activity relationships between
aggregations of the components into agents with properties. Zabezhailo et al
used expert consultations (from human experts) to encode these relationships.
Thus the method introduces the third level of the tri-level architecture proposed
above. In Zabezhailo's work, the methods of QAT were used to predict the
properties of assembled components based on SAR analysis. These methods use the
"outer meta-languages" of QAT as defined in Finn (1991) and Prueitt
(1998).
Quasi Axiomatic Theory (QAT)
is at the heart of the Russian control theory. The principle architect of QAT,
V. Finn, defines plausible inference as an inference that is not completely
reliable and where the set of falsifiers is not known completely. A reliable
inference is defined by Finn to be an inference that has no formal falsifiers.
Finn's system for reliable and plausible inference considers two cases; the
first where the set of falsifiers are generated, using specific procedures, in
a finite number of steps from the present logical description. This is the
"closed world" of a fixed formal system.
The second type of reasoning
occurs when it can be demonstrated that no effective procedure is known for
generating the full set of falsifiers for at least one relevant logical formula.
This case is an indication that the target of description is an open complex
system. A full range of citations is given in Finn (1991).
In most AI applications,
inferential falsification is formally defined rather than intuitively defined,
and thus we often face the difficult philosophical issues related to the
provability of a deductive sequence. The falsification of inductive inference
is not so easy, and the properties and causes of natural systems are not fully
understood, and change over time. Thus formalism supporting plausible reasoning
is necessary, as is proposed by Finn.
Natural situations are not
just a forced aggregation of a fixed store of basic elements into ensembles.
Likewise, complex models can not just be a grammatical aggregation of a fixed
set of primitives. Logical formulas may do quite well in simple cases; however,
for situational analysis to handle novelty and non-stationarity, a sequence of
measurements of the world at specific times and places must be made. This
suggests that a measurement of the environment occurs through a sequence of
perception-action cycles, each cycle composed of assembly, stability, and
disassembly phases. This central role of perception-action cycles links Russian
applied semiotics to American ecological physics.
Given the advanced state of
the several supporting schools of thought, we are very close to the big
paradigm shift that causes the nature of the discussion about knowledge
management to change. For example, during the perception-action cycle a number
of physical processes are known to take place; including pattern extraction,
memory management, iconic computation, behavioral responses, observations, and
the phenomena of cross scale entanglement of information.
However, very few software
systems support visualization of theme (feature) space as well as an integrated
representation of knowledge in the form of a concept database. Concept
databases do exist in a few cases (Abecker, et al, 1997), as discussed above, but
are primarily restricted to mechanical systems or industrial plants. Software
packages such as Spires makes projections into feature spaces based on cluster
analysis. But this technology is poorly received and poorly understood in the
market place. Thus, there is room for new innovations in software design and
implementation, based on the tri-level architecture.
The work is just beginning.
The scientific challenge is to see both computational visualization and human
perception within the same paradigm. This may not be so difficult. A large
amount of experimental research exists and some of this has been integrated
into explanatory frameworks such as the author’s General Framework for
Computational Intelligence (GFCI). The study of these frameworks suggests that
the human brain achieve the formation of concepts through a distributed
disassembly and reassembly of representational features.
The human concept space, in
vitro, may be pictured as a virtual space in the sense that the space does
not actually ever exist in total in any specific circumstance. In the GFCI,
parts of this virtual space come into being while various types of competitive
cooperative network dynamics block other parts. Of course, this simple
architecture disguises the complexity of how the brain uses both its neural
architecture and its chemical composition. The simplicity of the architecture
finds immediate rewards by providing a truly enterprise centric view of
organizations, a view that uniquely re-configures itself around each individual’s
location in a concept space.
A projection from a complete
enumeration of a knowledge engineering type concept space can be made onto a
concept subspace. This localization provides the promise of individual centric
knowledge processing within the sketched out outline of business and political
processes. The interpretation space may be mirrored by activation of components
of a situational model supporting automated reasoning. The mirrors can be
maintained by a neural network associative memory as demonstrated in (Dubchak
& Muchnik, 1995). The mirrors can be two way, and as a consequence a
separate mapping back to a concept subspace may be made after an inference
engine has changed the state space of the situational model. This notion of
projections and mirrors between representational spaces is a reasonable first
model for the production of computer based situational models within knowledge
processes about complex social systems.
Theoretical frameworks, such
as the GFCI, are developed to organize the experimental work. However, there
are a number of scientific issues that are not yet resolved.