U.S. patent application number 10/135159 was filed with the patent office on 2004-01-22 for adaptive dynamic personal modeling system and method.
Invention is credited to Goraya, Tanvir Y..
Application Number | 20040015906 10/135159 |
Document ID | / |
Family ID | 23103342 |
Filed Date | 2004-01-22 |
United States Patent
Application |
20040015906 |
Kind Code |
A1 |
Goraya, Tanvir Y. |
January 22, 2004 |
Adaptive dynamic personal modeling system and method
Abstract
The present invention provides a system and method for building
and analyzing a quantitative model from qualitative tacit
knowledge. In one aspect of the present invention, there is
provided a system for creating and analyzing a model of a user's
semantic knowledge. The semantic knowledge model is based on cause
and effect relationships as defined by the user. In another aspect
of the present invention, there is provided a system for creating
and analyzing a model of a user's episodic knowledge. The episodic
knowledge model is based on the user's past experiences, including
recalled stimuli and responses. The semantic and episodic models
are used to describe the users internal mental model.
Inventors: |
Goraya, Tanvir Y.;
(Bratenahl, OH) |
Correspondence
Address: |
TUCKER, ELLIS & WEST LLP
1150 HUNTINGTON BUILDING
925 EUCLID AVENUE
CLEVELAND
OH
44115-1475
US
|
Family ID: |
23103342 |
Appl. No.: |
10/135159 |
Filed: |
April 30, 2002 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60287536 |
Apr 30, 2001 |
|
|
|
Current U.S.
Class: |
717/141 ;
706/45 |
Current CPC
Class: |
G06N 5/022 20130101 |
Class at
Publication: |
717/141 ;
706/45 |
International
Class: |
G06F 009/45 |
Claims
What is claimed is:
1. A method for building and analyzing a quantitative model from
qualitative tacit knowledge comprising the steps of: receiving a
plurality of cause factors and a plurality of effect factors;
receiving a plurality of defined cause and effect relationships,
each defined cause and effect relationship comprising a
relationship value corresponding to a degree of influence that a
selected cause factor has on a selected effect factor; and
providing a means for holistically analyzing a plurality of cause
and effect relationships.
2. The method of claim 1, further comprising the step of applying a
linear function to the relationship value.
3. The method of claim 2, wherein the linear function is applied
based upon a specified relative importance.
4. A method for building and analyzing a quantitative model from
qualitative tacit knowledge comprising the steps of: receiving a
plurality of episodes, each episode comprising a plurality of cause
factors and a plurality of effect factors; receiving a plurality of
defined cause and effect relationships, each defined cause and
effect relationship comprising a value corresponding to a degree of
influence that a selected cause factor in an episode has on a
selected effect factor within the same episode; and matching
defined relationships in a first episode with defined relationships
in a second episode such that the matched relationships have the
same cause factor and the same effect factor; deriving a function
from the matched defined relationships such that the derived
function describes a relationship between causes and effects across
episodes; and providing a means for holistically analyzing a
plurality of cause and effect relationships.
5. The method of claim 4 wherein the derived function is a
nonlinear function.
6. A system for building and analyzing a quantitative model from
qualitative tacit knowledge comprising: computer readable code on a
computer readable medium for receiving a plurality of cause factors
and a plurality of effect factors; computer readable code on a
computer readable medium for receiving a plurality of defined cause
and effect relationships, each defined cause and effect
relationship comprising a relationship value corresponding to a
degree of influence that a selected cause factor has on a selected
effect factor; and computer readable code on a computer readable
medium for providing a means for holistically analyzing a plurality
of cause and effect relationships.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] The present application claims the benefit of U.S.
Provisional Patent Application Serial No. 60/287,536 filed Apr. 30,
2001.
BACKGROUND OF THE INVENTION
[0002] The present invention pertains generally to field of
artificial intelligence, and more specifically to a method and
system for adaptively and dynamically converting qualitative human
judgments into quantitative mental models.
[0003] Epistemology is the branch of philosophy that studies the
nature of knowledge, its presuppositions and foundations, and its
extent and validity. Within the epistemology community, there is no
generally accepted explanation for how tacit, experiential,
subconscious and qualitative knowledge, judgments, skills,
gut-feelings, and intuition actually work. There is, however, a
general consensus regarding the existence of two basic types of
knowledge: explicit and tacit.
[0004] While some epistemologists may define the terms "tacit
knowledge" and "intangible knowledge" differently, the two terms
are used interchangeably herein. Tacit knowledge can be defined as
the indescribable and intractable, yet commonplace knowledge that
people acquire and possess by virtue of their experience. It is
knowledge that is tempered by social and organizational norms, as
well as personal predilections. Therefore, it is dynamic and
evolves and refines over time. For example, a master craftsman is a
superior craftsman due to decades of experience and his modus
operandi is tempered by the norms of the environments to which he
has been exposed, reflecting his personal artistic tastes and
biases. His tacit, amorphous, or intangible knowledge is visible in
his work. Furthermore, he is able to make a wide variety of
decisions without being able to exactly explain the reasons for his
acts, much less provide precise rules.
[0005] However, an expert, such as a master craftsman, has no
mechanism for analyzing and interacting with his own tacit
knowledge. There is no method for engaging in a hermeneutic dialog
with his personal mental-model or intellectual text. Experts cannot
ask their personal mental-model to find solutions for constrained,
weighted goals. They have no techniques for reflecting upon their
knowledge-selves from various perspectives and combinations of
relative weights and constraints applied to their mental-models.
Unfortunately, creating a means for interacting with a tacit
knowledge mental-model is particularly difficult from a
practicality standpoint, and elusive from a theoretical standpoint.
In order to effectively interact with a tacit knowledge
mental-model, one might convert tacit knowledge into explicit
knowledge. However, from a theoretical standpoint, such a
conversion is problematic. If tacit knowledge is indeed tacit then
how can it be made explicit? Furthermore, if tacit knowledge can be
made explicit, is it tacit knowledge?
[0006] Different experts might perform identical tasks differently.
Sometimes, the same expert might perform the same task differently
at different times. Attempting to map tacit knowledge in itself is
an ambitious and ill-structured intractable problem. For instance,
even if one were to ignore the recondite nature of tacit knowledge
and computational limitations, an attempt to encode and execute
such voluminous knowledge is likely to cause a combinatorial
explosion. A combinatorial explosion can be illustrated by a
chessboard having sixty-four squares wherein a single grain of rice
is placed on the first square, two grains on the second square,
four grains of rice on the third square, and wherein the amount of
rice placed on each subsequent square is double the amount placed
on the previous square.
[0007] Unfortunately, a decision-maker is not necessarily aware of
how he arrives at a particular decision and of his own tacit
knowledge. Humans have biases and differ in how they process
information and make decisions. Intuitive decision-makers approach
a problem with multiple methods, using trial and error to find a
solution. It has been argued that choices are not made, but are
continuously being modified to accommodate changing objectives,
environments, value preferences and policy alternatives provided by
the decision-maker. Knowledge is created by cycling through active
experimentation and reflective observation.
[0008] There exists a need to capture an expert's tacit knowledge
that exists as complex cause and effect relationships and past
episodes. Most knowledge-based systems are concerned with the
explicable knowledge. Furthermore, existing tools for capturing
knowledge mainly comprise simple maps, lists, narrative, and causal
diagrams. Experts, in various walks of life, make decisions based
on judgment, gut-feel, and intuition, because they can draw upon
years of hands-on learning and past experiences. Such experts are
precious intellectual assets of an organization, and when they
leave, the organization is drained of decades of experience. To
date, there have been no computer-based methods to quantify and
retain this type of judgmental knowledge. Existing techniques, such
as rule-based expert systems, only retain brittle perceptions of
the experts, not their expertise. Also, they are very expensive and
require months or even years to build.
[0009] Typical Knowledge Process Management
[0010] While there is no satisfactory method to elicit tacit or
intangible knowledge, there are methods for extracting explicit
knowledge. Typically a knowledge engineer or a moderator interacts
with an expert. The knowledge engineer engages in a question and
answer session with the expert to elicit explicit knowledge. In
artificial intelligence ("AI"), the knowledge engineer might use
techniques that are suitable for a particular expert system or
programming environment. In a social science environment, the
moderator might use techniques such as analytical hierarchy process
("AHP"), cause-mapping, drawing on a surface, interviews, and
narrative, etc.
[0011] As a result of these knowledge acquisition techniques, the
knowledge engineer assembles a body of knowledge, which he
structures into a formal document in a form suitable for encoding
into a rule-based expert system or systems dynamics modeling
package. The analyst takes the documents from the knowledge
engineer and prepares system design documents in the form of data
flow diagrams ("DFD"), flowcharts, entity relationship diagrams
("ERD"), etc. The system design documents are then given to a
programming team, which codes the received information into a
software program. If the system is declared acceptable, then
potential users are trained on the software. In some instances,
however, the expert is not even involved in the use of the software
or does know how to use it. At this point, the knowledge
acquisition from the expert typically ends. However, if the system
is found to be unacceptable, the entire process must be repeated,
from the quest and answer session with the expert.
[0012] This method for knowledge process management is problematic
for a number of reasons. First, a disconnect exists between the
expert and the software program, or system. This disconnect creates
a situation where the knowledge engineer inadvertently adds his own
bias. In addition, the analyst and programming team add or remove
from the knowledge because of constraints such as system document
design, requirements or limitations of the programming environment,
or of the expert system shell. Also, because the system contains
only explicable knowledge, the expert may not be able to explicate
his know-how, and he may even be aware of the skills he has
acquired. In addition, different knowledge workers may approach the
same problem differently based on their personal sets of biases,
skills, and experiences. Additionally, because of the crisp
rules-based programming, the system does not allow interpolation
and is impossible to test comprehensively by firing all of the
rules due to combinatorial explosion. Furthermore, the system is
fairly static in that once built, implementing changes is
problematic. In other words, the system is inadequate because it is
costly, it does not permit direct interaction with the expert, it
does not contain tacit knowledge, it is strictly rule-based, and it
is not adaptable.
[0013] Existing Techniques
[0014] When viewed individually from the perspective of
constructing an environment to map cognition onto computation for
creating an adaptive interactive environment that captures
intangible knowledge, existing techniques have many shortcomings.
The primary reason is that these tools assume a rational
decision-making homoeconomicus and go to extra lengths to remove
human judgment from knowledge elicitation and model building
processes. In doing so, they remove the very ingredient necessary
for tacit knowledge.
[0015] On line Analytical Processing ("OLAP") based
decision-support systems are popular for storage, manipulation,
slicing and dicing, and presentation of data. These system include
variations of OLAP, such as relational OLAP ("ROLAP"),
multidimensional OLAP ("MOLAP"), hybrid OLAP ("HOLAP"), etc. and
can be used for representing and manipulating axiomatic human
knowledge in order to present the facts from multiple perspectives.
OLAP systems are problematic for a variety of reasons. First,
dimensions are pre-assigned by the architect, which implies
inflexibility in scaling the views. Also, in order to provide a
reasonable response time, less than 5 seconds, data has to be
precomputed for performance purposes. OLAP also suffers from the
curse of dimensionality. Published benchmarks of OLAP products show
a data explosion factor of 240, requiring 2.4 GB of storage to
manage 10 MB of input data. Finally, the heuristics or AI used in
OLAP are based on mathematical models which may or may not
correspond to actual usage patterns.
[0016] Probablistic systems both conventional and Bayesian, are
impractical for dynamic environments first because the sum of all
probabilities must equal one, and second because a priori
probabilities must be known in order for the system to work. In
order to counter this problem, some systems use frequency-based
probabilities. Such an approach negates the essence of probability
theory and renders it difficult to ascribe any reasonable degree of
certainty to the outcomes.
[0017] Bayesian Belief Networks are inspired by the work of
Reverend Thomas Bayes' philosophy and integrate a frequency-based
approach, based on probabilistic statistical theory, with a
tree-like structure. Bayesian Belief Networks share many
limitations and problems with Bayesian probability based systems in
a dynamic environment. The tree-like architecture and references to
frequencies of occurrence of events make it impractical for use as
a general-purpose solution.
[0018] Uncertainty theories are generally ad-hoc theories with
ad-hoc solutions. One example is the well-respected
Dempster-Schaffer Theory of Uncertainty. This theory, like many
others, basically relaxes the tenants of probability theory. In
addition, it combines it with set theory to incorporate and account
for uncertainty. The result is an extremely complex system that
requires a full-time mathematician or an AI expert for maintenance.
These systems are not adaptive because of their reliance on
up-front custom designs and become prohibitive because of their
high design and maintenance costs.
[0019] Statistics is one of the most important sciences of modem
times. Although we make many decisions based on statistics, it is
problematic when used for knowledge management. First, statistics
require Design of Experiments ("DOE"), a carefully controlled set
of results by running an experiment in a controlled environment at
specific data points in a problem space. This approach is not
possible in real-life dynamic situations. If the DOE is violated
then the confidence in statistical analysis cannot be held.
Secondly, when modeling a process, statistical analyses hold true
under strict conditions. For example, it is generally assumed that
relationships between objects are linear. This assumption is far
from true in the highly nonlinear real world. Statistics can model
moderately more complex, or nonlinear, relationships but at the
cost of accuracy and certainty. Third, statistics are not adaptive
in that they do not learn while maintaining existing knowledge.
Fourth, statistical solutions are static. That is, if there is a
change in any variables or if a variable is added or removed then
the entire results must be recomputed. This makes statistical
analyses inappropriate for dynamic environments.
[0020] Numerical vector analysis is based on a distance metric such
as the Minkowski Metric. The basic concept is that of a metric to
define distance or closeness that correspond to dissimilarities and
similarities. Euclidean Geometry is a special case of this metric
and is commonly used. The first step is to define a vector space
and then to populate it with data. Each data point is represented
as a vector in multi-dimensional space. The vectors, which lie
close to each other, are considered to be similar with a certain
degree of certainty. This is basic categorization of data. These
methods work well with numeric data but are of very little help
when dealing with text. The reason is that it is almost impossible
to define a distance metric between words. For example, what is the
distance between "cold" and "epistemology"? There are various
quasi-distance metrics for text that mostly use set theory.
[0021] Rule-based expert systems gained popularity during the
1980's. These systems are composed of rules provided by an expert
and a mechanism to invoke the rules using forward- or
backward-chaining algorithms. Knowledge acquisition for is a
bottleneck for such systems. It is time-intensive and iterative
human-intensive activity requiring systems analysis, interviewing,
and interpersonal skills. Experts may consciously or
sub-consciously have ulterior motives not to be forthcoming with
the entire or best information. Rule-based systems require domain
experts, considerable time, and knowledgeable engineers who have
familiarity with the domain. Experts may be too busy, difficult to
deal with, or unwilling to part with knowledge. There is not enough
understanding of how to represent commonsense knowledge. These
systems are not adaptive--they do not learn from mistakes.
Readjusting them can be a huge task. It is not easy to add
meaningful certainty factors on many rules. There can be
conflicting sources of expertise. These systems have low
fault-tolerance and exhibit inelegant degradation because they deal
with crisp rules.
[0022] Each of these factors considerably increases the cost of
building and maintenance. These systems are not adaptive, that is
they do not learn from previous mistakes. Furthermore, as the size
of a problem increases the number of rules increases, resulting in
high complexity, higher costs, and ultimately an unmanageable
system. Even addition of one more rule can have unexpected effects.
Therefore, these systems are useful when there are only a handful
of rules, when the problem domain is well defined in advance, where
the system is static, the knowledge engineers are unbiased and
experts are willing to part with their knowledge. Even when experts
into are willing to part with their knowledge they will only be
able to describe their explicit knowledge, not their tacit or
implicit knowledge, gut-feel, know-how, experiential knowledge, and
intuition.
[0023] Artificial Neural Networks ("ANN") are mathematical
representations of massive parallel processing systems loosely
modeled after the biological brain. During the 1990s these systems
saw a resurgence after almost three decades of quiescence. ANNs can
learn and discover relationships, or mathematical mappings, between
causes and effects from datasets. They can also be used for
mapping, optimization, and auto- and hetero-associative memories.
Common problems with ANNs are their inability to explain their
behavior--black box nature, overfitting of data, introduction of
noise, introduction of high degrees of nonlinearity, overtraining,
memory effect, difficulties with generalization or regularization.
Properly designing the architecture, preparing pristine data,
training, and interpreting the results requires an understanding of
mathamatics, probability, statistics, nonlinear multiple criteria
optimization, and the problem domain. The biggest strength of
neural nets--their ability to learn any relationship is perhaps
also one of their biggest shortcoming, since they can learn noise
and assume existence of improper relationships.
[0024] The present invention overcomes these problems by carefully
segregating pockets of dense data points from the sparse dataset,
by holding back data for regularization, by using optimization
algorithms for locating the best optimum in synaptic weight space,
and by parsimonious use of neurons.
[0025] Fuzzy Logic is an extension of classical logic. It is
concerned with approximate instead of exact reasoning. It allows
use of fractions instead of two discrete values. Fuzzy logic is an
inexact reasoning technique, which uses the concept of degrees of
membership in a set. This is an improvement because it allows the
use of such concepts as hot, warm, lukewarm, . . . , cold, etc.,
whereas classical logic could only allow for two concepts such as
hot and cold. Fuzzy systems are particularly successful in control
applications in industrial processes and consumer goods as embedded
components. An example of its success is the braking system of
Bullet Trains. Fuzzy logic has only been successfully applied in
systems with a small number of variables. Fuzzy logic requires a
great deal of knowledge of the problem domain. The number of rules
can grow exponentially with the number of variables and number of
possible choices, rendering even a moderate problem incapable of
implementation.
[0026] Success can be achieved by building and extensively
fine-tuning each particular fuzzy logic system specific to a
well-defined task. Fuzzy logic can provide model-free information.
However, fuzzy logic has not yielded tools to easily convert
qualitative information into a robust quantitative model. Fuzzy
systems are not adaptive. Constructing and validating a rule base
is an iterative and difficult task. Experts and time are needed to
design, construct, validate and fine-tune fuzzy systems. This
includes manually tuning membership functions and fuzzy rules.
Fuzzy systems become very expensive as size and complexity of the
problem increases. It is not easy to prove that during operation a
fuzzy system will remain stable and not become chaotic because it
can be very sensitive outside its operating range. There are also
mathematical criticisms of fuzzy logic that under certain
conditions fuzzy logic collapses to two-value logic--the
paradoxical success of fuzzy logic. Fuzzy systems also lack audit
trails.
[0027] Another major difficulty is in defuzzification. There are no
consensus methods for automatically applying defuzzification. This
means that each system has to be carefully handcrafted for a
particular problem and it is neither scalable nor adaptive. This
ultimately increases both the computational and design costs. The
task of defuzzification is not only quite complex but also
controversial. Generally, iterative trial and error and
approximations are required to defuzzify results into a meaningful
and acceptable form. Defuzzification alone makes the idea of fuzzy
logic based tools unsuitable for building robust generalizable
models.
[0028] Fuzzy logic has achieved modest success in model building in
concert with other model building methods such as expert systems.
But even in hybrid systems many of the above mentioned difficulties
remain. Some of the more successful methods have combined neural
nets with fuzzy logic to address some of these problems, e.g.,
Adaptive Network Fuzzy Inference System ("ANFIS").
[0029] Judgmental Bootstrapping ("bootstrapping") is a method used
in forecasting to create quantitative models from human judgments.
There are two types of judgmental bootstrapping: direct and
indirect. The primary difference between bootstrapping and other
expert systems is that for bootstrapping, experts are not asked to
furnish rules; instead rules are "inferred" from their behavior. A
quantitative model of an expert's rules is constructed by observing
his judgments while he makes forecasts. Generally expensive
protocol studies are required to gather this data. The causal
variables used by the expert are treated as independent variables
and his forecasts as dependent variables. From this observed data,
ordinary least square regression is used to formulate linear
quantitative relationships between the forecasts and the causal
variables. The resulting mathematical equations or models represent
the expert's rules.
[0030] Bootstrapping has some limitations. It requires the delicate
task of observing experts during the performance of their tasks and
translating the observations into numeric form. This raises
fundamental socio-behavioral issues such as the Hawthorne effect,
observational observations and dual interpretation process of
distinguishing between "act meaning" (meaning of act to the actor)
and "action meaning" (meaning of act as observer's subject matter).
Bootstrapping models are not adaptive. These models, which are
often supposed to aid experts, are not easily modifiable especially
by the experts. Modifications may be necessary when either the
expert's mental model or the external environment have changed. The
model needs to be reformulated from scratch when many changes occur
such as addition of new causal variables.
[0031] Extrapolation is also a concern. Reported studies have used
cross-sectional data instead of time series data. The major
problem, which is lack of adaptation remains even if experts are
involved in validating and using their models. When modifying
models, usually existing models have to be abandoned and new models
need to be formulated from ground zero. Although much less data are
required than, for example, conjoint analysis, at least five to ten
experts are typically needed for useful bootstrapping models. In
addition, since regression is used to build the model all the
shortcomings of statistical regression also influence
bootstrapping.
[0032] Analytic hierarchy process ("AHP") is a "multicriteria
decision model that uses hierarchic or network structure to
represent a decision problem and then develop priorities based on
decision-maker's judgments throughout the system. AHP has been
applied with success in areas where judgment is to be taken into
consideration such as facility location analysis, work scheduling,
planning, capital budgeting and technology transfer. One advantage
of AHP is its ability to measure consistency. The most important
limitation of AHP is exhibited when the ranks determined for
alternatives through the AHP change as new alternatives are added.
Dyer proposed that this problem could be addressed if weights were
expressed using an interval scale as opposed to the ratio scale.
This requires borrowing concepts from multiattribute utility
theory. Some have opposed this method of correction, arguing that
AHP is a standalone body of research not an extension of
multiattribute utility theory.
[0033] Some other technologies worth mentioning include Systems
Dynamics Modeling, Simulated Annealing, and Markov and Hidden
Markov Processes. None of these provide adequate means for mapping
cognition onto computation and its inverse.
BRIEF SUMMARY OF THE INVENTION
[0034] It is therefore an object of the present invention to
provide a method and system for building and analyzing a
quantitative model from qualitative tacit knowledge.
[0035] In accordance with the present invention, there is provided
a method for building and analyzing a quantitative model from
qualitative tacit knowledge. According to the method, a plurality
of cause factors and a plurality of effect factors are received. A
plurality of a plurality of defined cause and effect relationships
are also received, wherein each defined cause and effect
relationship comprises a relationship value corresponding to a
degree of influence that a selected cause factor has on a selected
effect factor. A means for holistically analyzing the plurality of
cause and effect relationships is provided.
[0036] Also in accordance with the present invention, there is
provided a method for building and analyzing a quantitative model
from qualitative tacit knowledge. According to the method, a
plurality of episodes are received, wherein each episode comprises
a plurality of cause factors and a plurality of effect factors. A
plurality of defined cause and effect relationships are also
received, wherein each defined cause and effect relationship
comprises a value corresponding to a degree of influence that a
selected cause factor in an episode has on a selected effect factor
within the same episode. Defined relationships in a first episode
are then matched with defined relationships in a second episode
such that the matched relationships have the same cause factor and
the same effect factor. A function is then derived from the matched
defined relationships such that the derived function describes a
relationship between causes and effects across episodes. A means
for holistically analyzing the plurality of cause and effect
relationships is provided.
BRIEF DESCRIPTION OF THE DRAWINGS
[0037] FIG. 1 represents the overall conceptual framework for the
present invention;
[0038] FIG. 2 is a flow chart representing the system functionality
in response to a user's hermeneutic tacking back and forth
according to the present invention;
[0039] FIG. 3 is representation of a stimulus and response spaces
according to the present invention;
[0040] FIG. 4 is a screen shot of the e-model educer screen
according to the present invention;
[0041] FIG. 5 is representation of the mapping and consultation
process of the e-model according to the present invention;
[0042] FIG. 6 is a dendogram representing the results of
categorization of episodes and factors according to the present
invention;
[0043] FIG. 7 is a screen shot of a two-dimensional plot of the
type generated by the e-model upon selection of a single
stimulus;
[0044] FIG. 8 is a screen shot of a three-dimensional surface plot
of the type generated by the e-model upon selection of multiple
stimuli;
[0045] FIG. 9 illustrates an output screen showing general
statistics;
[0046] FIG. 10 illustrates a graph showing the relationship of all
stimuli to a single response factor, when all other response
factors were held at their mean values;
[0047] FIG. 11 illustrates a control console screen for learning
the mapping in the e-model;
[0048] FIG. 12 illustrates the overall design of the e-model
according to a presently preferred embodiment;
[0049] FIG. 13 is a high level diagram illustrating some of the
functionality of the e-model;
[0050] FIG. 14 is a screenshot illustrating a main control panel
for s-modeling, clustering, ranking, goal seeking, and
visualization according to the present invention;
[0051] FIG. 15 is a connectionist representation of the
architecture for computing a j.sup.th effect for computation of the
s-model according to the present invention;
[0052] FIG. 16 is a simplified connectionist representation of FIG.
15;
[0053] FIG. 17 is a screen shot illustrating consultation results
and a faceted decision hyper-plane surface of the s-model of the
present invention;
[0054] FIG. 18 screen shot illustrating two different manipulable
graphical views of the relationships and the results of corner
solutions optimization;
[0055] FIG. 19 is a screen shot illustrating an interactive
graphical display of the relationships and the adaptive conjoint
feedback mechanism for cybernetic control or teaching the system
graphically by pushing and pulling on the graph;
[0056] FIG. 20 is a screen shot illustrating cascaded windows where
each window shows relationship vectors for all causes in one
effect's space;
[0057] FIG. 21 illustrates the overall design of the s-model
according to a presently preferred embodiment;
[0058] FIG. 22 is a high level diagram illustrating some of the
functionality of the s-model.
[0059] FIG. 23 is a system diagram of the present invention in a
network environment; and
[0060] FIG. 24 is an illustration of the structure of the e-model
and s-model according to a presently preferred embodiment.
DETAILED DESCRIPTION OF THE INVENTION
[0061] The present invention provides a system and method for
capturing and preserving one's tacit knowledge. In doing so, it
provides a unique computer environment, based upon advances in
artificial intelligence, which captures and retains tacit
judgmental type knowledge in a computer and transforms sparse
qualitative data into mathematical models.
[0062] Turning now to FIG. 1, the overall conceptual framework for
the present invention is provided. The framework suitably comprises
three basic environments: an expert environment 100, a computing
environment 102, and an interaction environment 104 for
facilitating interaction between human and machine. Within the
expert environment 100, there exists an expert's mental model 106,
which suitably comprises intuition 108, which is a tacit,
non-explicit type of knowledge. Intuition can also be described as
intangible amorphous knowledge. An expert's mental model 106 also
comprises episodic memory 110, semantic memory 112, and experience
114. An expert utilizes his mental model 106 to exercise judgment
116.
[0063] The expert's judgment 116 is used to create a
moderate-resolution model 118 in the computing environment 102. In
the computing environment 102 an expert uses his judgment 116 to
creates and then maintain a moderate-resolution model 118, which is
a computer-based rendition of his mental-model. The
moderate-resolution model 118 is preferably an aggregate of least
two models: a semantic model and an episodic model.
[0064] In the human-machine environment 104, an expert interacts
with the model 118 at process block 120. The expert suitably
interacts with the model 118 by asking questions, discovering
categories, assigning relative weights, assigning costs to goals,
and performing goal seeking. During interaction at process block
120, the expert suitably performs categorization, what-if analyses,
and/or multi-criteria optimization. The mechanics of the underlying
computational machines are preferably hidden from the expert such
that the expert sees only easy-to-use controls. The results are
suitably provided via visual cognitive feedback at process block
122. The results of his queries are preferably presented to him in
easy-to-interpret and manipulable colorful graphics. The expert is
preferably able to create queries out of the results, thus engaging
in a cyclical hermeneutic type interaction.
[0065] Progression then continues to process block 124 wherein the
expert evaluates the results obtained from interacting with the
system in light of his judgment 116 mental model 106. Flow then
progresses to decision block 126 wherein a determination is made
whether the model 118 coincides with the expert's mental model 106.
A positive determination at decision block 126 causes progression
to process block 130 wherein the expert suitably engages in an
adjustment of his mental model 106 based on the results returned by
the model 118. In other words, the expert is enlightened as to his
mental model 106 through analysis of the results returned.
[0066] A negative determination at decision block 128 causes
progression to process block 128 wherein the parameters used to
create model 118 are adjusted. In the presently preferred
embodiment, the invention comprises an Adaptive Conjunctive
Feedback Mechanism ("ACFM"), which provides the expert with the
capability to modify the model 118 upon evaluation of the results
such that the model more closely matches his mental model 106. If
he feels that the result is incorrect then he can click and drag on
the result's graphical representations to bring them to the values
that he thinks ought to be. The ACFM then automatically adjusted
the model 118 such that the model learns the changes applied by the
expert. In addition to teaching or correcting a model, these
features also allow for adaptation of the model as the expert
grows.
[0067] Turning now to FIG. 2, a diagram representing the system
functionality in response to a user's hermeneutic tacking back and
forth according to the present invention is provided. The basic
flow commences at process block 200 wherein models, such as model
118 are constructed and simulated. Progression then flows to
process block 202 wherein sub-models are constructed. In other
words, model sub layers are suitably constructed. Such construction
is suitably accomplished through the use of a plurality of engines,
which are computer readable code, and are suitably software
components. In the presently preferred embodiment, such engines
include a Self Adaptation System; a mapping system that is
preferably a non-linear mapping system; an episodic memory system;
an Optimization/Goal Seeking System; an Adaptive Clustering System;
a Ranking System; an ACFM; and a Plotting/Graphing Engine.
[0068] Progression continues to process block 204 wherein
mathematical instructions and data representative of the
constructed models and sub-models are generated. Flow then
continues to process block 206 wherein the generated mathematical
instructions and data are translated into user interface ("UI")
information. This translation is suitably accomplished by any
method as will be appreciated by those skilled in the art.
Preferably, the mathematical instructions are translated into
graphical user interface ("GUI") information.
[0069] Progression then flows to process block 208 wherein the
generated UI information is presented to the user. In the presently
preferred embodiment, GUI information is presented to a user in the
form of graphs, such as three-dimensional graphs, clusters showing
relative rankings, results of goal seeking, etc. In addition, the
presented GUI information is preferably interactive. Flow then
continues to process block 210 wherein a user response is received.
The GUI is preferably equipped with user-friendly controls, such as
sliders, buttons, checkboxes, etc.
[0070] Preferably, the user response is received from a keyboard,
mouse, touch screen, or other input device as will be appreciated
by those skilled in the art. Progression then continues to process
block 212 wherein maps are created based on responses received from
the user. The maps are suitably high-level maps that facilitate
user interaction through I/O devices. Following creation of
high-level maps and user interaction, progression flows to process
block 214 wherein the user interface signals received during
interaction with the user interface are translated into low-level
mathematical instructions and data. Flow then loops back to process
block 200 wherein the mathematical instructions and data are used
to construct and simulate models.
[0071] The present invention comprises a plurality of systems and
engines. First, the system comprises at least two memory based
modeling systems: a semantic model ("s-model") and an episodic
model ("e-model"). Additional models are also suitably included in
the system. The s-model is created from the supposed mapping of the
semantic memory or cause and effect relationships that an expert is
capable of expressing. In the presently preferred embodiment, the
s-model is a linear model that utilizes parallel distributed
processing representation. The e-model is constructed from past
experiences of an expert and is designed to mimic episodic memory.
In the presently preferred embodiment, the e-model is a nonlinear
model that utilizes parallel distributed processing
representation.
[0072] The present invention also comprises an interactive adaptive
categorization system ("ACS"). The ACS allows entered data to be
categorized and viewed from various perspectives. In the presently
preferred embodiment, the ACS is a hierarchical, agglomerative,
adaptive, neural network representation of computational machinery
that uses the Minkowski metric of second order or Euclidian
distance metric for numeric categorization. An expert can engage in
a dialog with the categorization engine through a user interface
and can increase or decrease a similarity measure to a desired
level. The categorization results are preferably presented in color
and as dendrograms.
[0073] The present invention further comprises a plurality of
optimization systems. Optimization is the selection of the best
item or combination of items from a group of many items or
combinations thereof. The presently preferred embodiment is capable
of performing a plurality of optimization methods. The system is
preferably capable of performing optimization by the Adaptive
Flexible Tolerance method, which is based upon the flexible
tolerance method, also known as the amoeba search algorithm, as
will be appreciated by those skilled in the art. The present
invention also comprises Evolutionary Programming optimization
functionality, which employs method that uses genetic algorithms.
The Adaptive Flexible Tolerance and Evolutionary Programming
optimization methods are particularly useful for optimization with
nonlinear models. The present invention further comprises Comer
Solutions optimization functionality, which employs a simple search
method that searches for best solutions at corners of a search
space. The Comer Solutions optimization method is particularly
useful for optimization with linear models.
[0074] In the preferred embodiment, the basic algorithms are
enhanced to include constraints and relative weights and are
capable of learning from their experiences as they optimize
different problems. While the enhancements do allow the system to
make generalizations for all types of problems since each problem
has a different solution space, they do improve results when
searches are conducted on the same or similar solution spaces. This
also enhances their ability to overcome local optima as more
searches are carried out on the same or modified versions of a
solution space. Preferably, the algorithms are capable of
dynamically adjusting their parameters as a search continues. This
functionality is particularly useful when an expert does not have
any knowledge of the underlying computational machinery, which is
the case in the presently preferred embodiment. If an expert is
knowledgeable in the field of computational intelligence, the
system preferably comprises a parameters screen whereby the expert
can enter parameters, pause an optimization, modify parameters, and
explore and study the results. In all cases, the preferred
embodiment of the present invention comprises an output in color
graphical format. In addition, an expert preferably has the
capability of retrieving numeric information from the various
underlying mathematical models.
[0075] The present invention also preferably comprises a visual
cognitive feedback system. Because more processing area in the
brain is dedicated to visual information processing than to any
other sense, there is extensive use of visual cognitive feedback.
Results are preferably shown in two- and three-dimensional
graphical forms with color as the fourth dimension.
[0076] E-Model
[0077] The purpose of the e-model is to allow an expert to create,
manipulate, and engage in a dialogue with a computer-based
rendition of his experiences. The basic premise behind the e-model
is that an episode, schema, or frame can be expressed as two
vectors, which in turn can be represented as relative numeric
values for stimulus (cause factor) and response (effect factor)
variables.
[0078] Turning now to FIG. 3, a representation of a stimulus and
response spaces according to the present invention is provided.
Mathematically, the system is suitably depicted as two spaces, a
stimulus space 300 and a response space 302, and stimulus vectors
304, and response vectors 306. The stimulus vectors 304 are
suitably representative of a one-to-one relationship and the
response vectors 306 are suitably representative of a one-to-many
relationship. The dimensionality of the stimulus and response
vector spaces preferably corresponds to the input and output
dimensionality of the e-model. The objective is to build an
injunctive mapping and an inverse mapping between the two spaces to
discover functional relationships.
[0079] Neural networks are suitably used to learn such a mapping,
but one must be careful to avoid over learning (memory effect). In
addition, one must be sure to allow for regularization, and avoid
unnecessary nonlinearities and noise. One less preferred method of
building such a mapping is to train a neural network. In the
presently preferred embodiment, however, patterns are first
separated into clusters using a similarity measure, and then neural
network models are built for each group of similar patterns. This
preferred method not only identifies outliers but also results in
smoother learned functions. These piece-wise functions are then
preferably combined to portray the overall learned functional
hyper-surface. These internal computations are suitably performed
without the knowledge of the user and within acceptable response
time for preparing the visual cognitive feedback for the user. The
time between the user submitting his request and the system's
response is suitably less than ten (10) seconds and most preferably
less than five (5) seconds so that a sense of continuity is
maintained.
[0080] Experts with an advanced working knowledge of computational
intelligence can modify various underlying parameters to achieve
desired mapping quality through an optional parameters screen. The
system preferably shows the mapping phases with graphical views,
including bird's eye and zoomed-in views of epoch learning errors
for each mapping. The epoch learning error is preferably shown with
high precision scientific notation. The system suitably permits
advanced users to pause, modify factors, resume, or abort a
learning session.
[0081] Turning now to FIG. 4, a screen shot of the e-model educer
screen according to the present invention is provided. After an
expert specifies the number of stimulus and response variables, the
expert suitably utilizes the educer screen to construct a
representation of an episode. For each specified stimulus, there
preferably exists an input-dimensionality slider bar 400. For each
specified response, there preferably exists an output
dimensionality slider bar 402. The slider bars 400 and 402 are
preferably pre-programmed for small and large increments. This
suitably enables the expert to fine-tune an episode by implementing
either fine or coarse changes. The expert suitably moves the
sliders 400 and 402 to construct the representation of an episode.
Through the educer screen, the expert suitably assigns labels to
stimuli, responses, and episodes suitably adds, deletes, or
modifies episodes. In addition, the episodes are optionally tagged
with a textual description by the expert to aid in recall and
recognition.
[0082] Turning now to FIG. 5 the mapping and consultation process
of the e-model is provided. The stimulus vector inputs 502 are sent
through a filtering or categorization system 504. The filtering and
categorization system suitably associates similar episodes in the
form of clusters while flagging any outliers. The system preferably
identifies and groups episodes that are similar to each other using
Minkowski metric. The filtering and categorization is suitably
accomplished through the utilization of statistical and autonomous
parallel processing systems. Each cluster from a hierarchy of
clusters built during categorization suitably feeds a neural net
system 506 comprising a stimulus space and a response space. It
should be noted that each neural net system 506 suitably comprises
more than one neural network and is not limited to a single
stimulus space or a single response space. A nonlinear injunctive
mapping is preferably learnt between stimuli and responses. For
each neural net mapping, the system preferably tests a plurality of
mapping techniques and chooses to utilize those mappings that
provide a balance between regularization, under-fitting, and
excessive nonlinearity. The system preferably automatically solves
for the constrained multi-criteria optimization problem each
model.
[0083] The model building process suitably comprises many steps
that are preferably hidden from the user. The results of consulting
the mapping are synthesized and are presented to the user as visual
cognitive feedback. After the mapping process is successfully
completed the expert can move through the original episodes. As he
moves through each episode, the values for stimulus and response
sliders are preferably automatically set to the values of that
particular episode.
[0084] The system also preferably performs inverse mapping, which
is the mapping from response space to stimulus space, in order to
seek out optimum stimulus values for desired response values.
Inverse mapping aids in optimization and goal seeking
functionality. As shown in FIG. 3, one-to-many mappings are
possible when mapping from response space to stimulus space.
Therefore, different combinations of stimuli can produce identical
responses. Simple inverse mapping causes neural networks to learn
the average of all the stimuli and associate that average with the
response. If different stimuli have different relative weights then
the system suitably associates a weighted average with the
response.
[0085] Rather than implementing simple inverse mapping techniques,
the present invention preferably implements an outside optimization
loop. In the presently preferred embodiment, a direct optimization
technique is utilized. A direct optimization technique is one that
evaluates the value of an objective function at various points in a
solution space, rather than using information form gradients,
conjugate gradients, and the like. Therefore, when using a direct
optimization technique, information conjugate gradients and the
like is not available. The optimization algorithm thus becomes more
important when conducting efficient searches. The presently
preferred embodiment employs two optimization techniques, each of
which includes measures for accommodating constraints, costs, and
relative desirabilities of the goals specified by the user. The two
optimization techniques are modified forms of an evolutionary
programming technique and a flexible tolerance or amoeba
method.
[0086] Ideally, one of the goals of the system is to reduce the
number of evaluations of an objective or merit function, while
achieving relatively quick evaluation times. Therefore, heuristics
are employed. One method that is suitably used to accomplish this
goal is to continually learn about the surface of the solution
space during the searching process. A short-term memory is suitably
built during optimization for a particular problem. This short-term
memory is preferably adaptive and capable of improving itself for
use with similar problems. As more searches are conducted, without
many changes in the parameters, this memory preferably begins to
aid in the location of the global optimum and helps avoid getting
trapped in local optima. It also helps in selection of
progressively better starting points as searches continue. Using
this configuration, the objective function is evaluated repeatedly
and the search continually progresses toward better solutions.
[0087] At any given time, a list of the best solutions obtained
thus far is preferably maintained. This is to ensure that if the
search is terminated, or if it wanders off in a wrong direction, or
if it takes a long time, or if the user is not satisfied with the
optimum solution, the user will still have a list of some
potentially viable options. The search suitably continues until a
pre-specified number of iterations is reached, a pre-specified
amount of time is elapsed, or a solution is found within a
pre-specified tolerance of the goals, weights, and constraints
desired by the user. At all times during the optimization, the user
suitably has the capability to pause a search, modify parameters or
his goals, restart the search, or stop the search.
[0088] The present invention also preferably provides functionality
that permits a user to recall past episodes from a repository of
episodes, by supplying weighted, partial, distorted, or full cues.
A search conducted on a partial cue suitably reduces the
dimensionality of the space to that of the partial cue whereas a
search based on a full cue suitably locates the closest matches in
Minkowski space. The user can thus perform a search to favor
dimensions that are weighted higher or lower. In addition, the
present invention preferably permits a user to search for a past
situation that is similar to a present situation.
[0089] The expert can consult the e-model by setting the slider
bars on the stimuli side to desired values and asking the e-model
to provide the results by consulting the model. Internally it is a
two-stage process of consulting the groups and then the associated
mapping. The values of the responses are obtained through
interpolation. A well-behaved model is one that allows
generalizations while minimizing noise, high nonlinearities, and
memory effect.
[0090] In addition to setting the sliders, the results of
consultation also appear on the same graphic display as the results
of consultation from the s-model. This makes it possible to readily
compare and contrast the results from the two different modeling
approaches, or espoused theories with theories-in-use.
[0091] The e-model also preferably allows the user to perform
grouping, categorization, or clustering of stimuli, responses, and
episodes. The user can graphically instruct the system to use a
selected similarity measure. In Euclidean distance terminology, a
cluster would be a circle in two-dimensions, a sphere in
three-dimensions, and hyper-sphere in higher dimensions. Clustering
is suitably repeated until either no factors (stimului and
responses) switch clusters or the number of factors switching
clusters is small, or the changing of clusters follows a
pattern.
[0092] Turning now to FIG. 6, a dendogram representing the results
of categorization of episodes and factors is provided. The results
of clustering are preferably presented in color such that similar
factors are represented by identical colors. The dendrograms
preferably represent the overall hierarchical structure of
categories and subcategories showing where various factors fit. The
dimensionality of the categorization space is preferably equal to
the number of stimulus and response variables in the case of
episodes, and equal to the number of episodes in the case of
stimulus and response categorization.
[0093] The user can also view the trends learned or discovered
during the mapping phase by selecting a stimulus or stimuli and
plotting a selected response as a function of the selected stimulus
or stimuli. When a user selects a single stimulus, the system
preferably generates a two-dimensional line plot of the type
illustrated in FIG. 7. Preferably, the line plot is depicted in
color. The curvature, slope, and direction of the plot show trends.
In addition, the system calculates an approximate slope and
presents it to the user to compare it with a relationship value
between the corresponding cause and effect variables of the type
provided by the user in the s-model. A plot is suitably generated
by holding all values constant except the selected stimulus, and
varying the independent variables at regular intervals as specified
by the user. A higher number of steps or intervals provides fine or
high-resolution plots. In order to generate the plot, the system
suitably consults the e-model a number of times equal to the number
of intervals.
[0094] When a user selects a plurality of stimuli, the system
preferably generates a three-dimensional surface plot of the type
illustrated in FIG. 8. The three-dimensional surface plot is
preferably color-coded to provide a better view and understanding
of the variables height that depicts the dependent variable. The
expert suitably selects the number of steps for which the system
should generate values for plotting. A plot is suitably generated
by holding all values constant except the selected stimuli, and
varying the independent variables at regular intervals as specified
by the user. A higher number of steps or intervals provides fine or
high-resolution plots. In order to generate the plot, the system
suitably consults the e-model a number of times equal to number of
intervals squared. In the presently preferred embodiment, the user
can rotate the plots around all three axes, change perspective,
tilt, and zoom in and out. These features facilitate the detection
of minor hidden anomalies in an otherwise apparently well-behaved
relationship trend.
[0095] In addition to the dendogram and discovered relationship
plots of FIGS. 6-8, the preferred embodiment of the present
invention generates additional output shown in FIGS. 9-11.
[0096] FIG. 9 illustrates an output screen showing general
statistics.
[0097] FIG. 10 illustrates a graph showing the relationship of all
stimuli to a single response factor, when all other response
factors were held at their mean values.
[0098] FIG. 11 illustrates a control console screen for learning
the mapping in the e-model.
[0099] FIG. 12 illustrates the overall design of the e-model
according to a presently preferred embodiment.
[0100] FIG. 13 is a high level diagram illustrating some of the
functionality of the e-model.
[0101] S-Model
[0102] The purpose of the s-model is to allow an expert to create,
manipulate, and engage in a dialogue with a computer-based
rendition of the factual relationships that he is capable of
articulating. The s-model is based on the premise that it is
possible to elicit piece-wise linear functional relationships
between pairs of causes and effects, or antecedents and
consequents, from an expert in order to construct a reasonably
plausible faceted decision-surface by joining the individual lines.
The input dimensionality of the model thus formed equals the number
of causes and the output dimensionality is equal to the number of
effects. The presently preferred embodiment implements the s-model
using a parallel distributed processing architecture.
[0103] The concepts of the basic elements of this technique are
similar to that of judgmental bootstrapping. However, bootstrapping
is a static technique whereas the s-model is part of a dynamic and
adaptive environment that is controlled by the expert. The s-model
is a tangible model that is suitably queried, whose elements can be
categorized, and whose factors can be weighted against each other.
It suitably provides visual cognitive feedback to the user, allows
for changing of the model by moving a few graphical artifacts, and
provides functionality for performing constrained, weighted, goal
seeking.
[0104] In order to more clearly explain the functionality of the
s-model, the following vocabulary will be used:
[0105] I The total number of causes;
[0106] J The total number of effects;
[0107] i The i.sup.th cause;
[0108] j the j.sup.th effect;
[0109] w.sub.ij Judgment-based value of the relationship between
the i.sup.th cause and the j.sup.th effect wherein the value is
assigned a real number between -10 and +10, the positive or
negative sign corresponding to positive or negative effect
respectively and the number representing the strength of the
relationship;
[0110] x.sub.i Value of the cue for the i.sup.th cause presented by
the user to the system for consulting;
[0111] r.sub.i Relative importance of the i.sup.th cause with
respect to the rest of the I-1 causes, wherein the default is
preferably set to r.sub.i=1/I, where all causes have equal
importance;
[0112] b.sub.j Expert user's bias or predilection towards or
against an effect, wherein it is a measure of relative importance
of the j.sup.th effect, as viewed by the expert, in the model;
[0113] o.sub.j Model's output value of the j.sup.th effect which is
the result produced by the model after consulting with a set of
causes, x;
[0114] f.sub.ij Expert's faith in the relationship between the
i.sup.th cause and j.sup.th effect wherein the relationship is
specified by w.sub.ij;
[0115] d.sub.j Desired value of the j.sup.th effect, wherein the
result produced by the model after consulting with a set of causes,
x, is o.sub.j, and wherein the user indicates that he prefers the
value d.sub.j instead of o.sub.j and would like the model to adapt
to reflect this change in his viewpoint;
[0116] D.sub.j The difference between the desired value and the
model's output for the j.sup.th effect,
[0117] D.sub.j=(d.sub.j-o.sub.j);
[0118] .tau. Similarity measure, or tolerance, a scalar used for
classification, wherein its value is directly proportional to the
degree of desired similarity;
[0119] Turning now to FIG. 14, a screenshot illustrating a main
control panel for s-modeling, consulting, clustering, ranking, goal
seeking, and visualization is provided. To initiate the s-modeling
process, a user suitably enters a plurality of names of the causes
(cause factors) 1402 or inputs for the model as row labels and a
plurality of names of effects (effect factors) 1404 or outputs of
the model as column labels. An empty matrix is created thereby
created. Each cell of the matrix corresponds to a judgment-based
qualitative relationship between a cause factor and an effect
factor.
[0120] For each cell, a user enters a number, w.sub.ij, between -10
and +10 to indicate the strength and direction of the relationship.
In other words, the cell value, w.sub.ij, 1406 represents the
influence that the corresponding cause has on the corresponding
effect. Optionally, the user also enters for each cell a number,
f.sub.ij, between 0 and +10 to indicate the expert's faith in his
judgment about the relationship. The system default is suitably
equal faith for all relationships. Also optionally, the user
suitably indicates a value, r.sub.i, for a cause or a plurality of
causes, its importance relative to the other causes. The user
suitably selects r.sub.i values through the use of slider bars
1408. The system default is suitably equal importance for all
causes. Furthermore, for each column or effect, a user optionally
indicates his bias or predilection toward or against the effect by
indicating a value, b.sub.j. The system default is suitably no
bias. If there is any special case resulting from atypical
conjunctions of certain causes, the user can add an appropriate row
label for the case.
[0121] Once the s-model is built, a user can validate the model. To
validate the model, the user suitably presents a set of values for
causes, consults the model, and observes the results produced by
the model. If the resulting effects produced by the model are not
satisfactory, the user suitably makes adjustments to the previously
entered values. If the results after modification are satisfactory,
the validation is complete.
[0122] The s-model preferably permits users to change values of up
to five parameters and view the resulting clusters. It can aid the
user in exploring different options and scenarios. It also provides
an understanding of similarities among causes and effects from
various viewpoints.
[0123] The system preferably automatically classifies causes on the
basis of similarity. A user suitably selects a similarity measure
1410. The presently preferred embodiment color codes the cause
factors 1402 such that similar factors 1402 are shown in the same
color. Similarly, the presently preferred embodiment color codes
the effect factors 1404 such that similar factors 1404 are shown in
the same color. In addition, color dendrograms of the type shown in
FIG. 6 show hierarchical clustering of causes and effects. Whenever
a change is made to a judgment value, faith value, the relative
importance of the causes, or to the biases towards the effects, the
system preferably automatically performs reclassification.
Immediate visual feedback is provided to the user through changing
colors, and new trees, to signify similarities among causes and
similarities among effects.
[0124] The expert can also increase or decrease the similarity
measure 1410 or tolerance, .tau.. Preferably, this is accomplished
either visually through a spin button or by typing a numeric value.
The tolerance is suitably graphically shown as a growing or
shrinking filled circle 1412. Immediate reclassification and the
resulting color feedback takes place whenever the similarity
measure 1410 is modified. Low and high tolerance values suitably
correspond to fine and coarse grains of classification
respectively.
[0125] The user can consult the model for what-if analysis by
presenting different values 1406. A user can provide qualitative
information about causes on continua that range from nothing to
extremely high. Alternatively, the user can also enter crisp
numbers. The system visually displays the corresponding effects
after consulting the model.
[0126] In addition, during goal seeking, a user can move goal
seeking sliders 1414 to desired values of causes and consult the
results of the model. Also, goal seeking operations suitably set
the slider bars 1414 to their appropriate positions. Furthermore, a
user can select relative desirabilities 1416 during goal seeking or
select which causes will be manipulated during goal seeking 1418.
Goals for causes 1420 during goal seeking are also suitably set to
either increase, decrease, or don't care.
[0127] When a set of causes x.sub.i, i=1, . . . , I is presented to
the s-model for consultation, the outputs o.sub.j; j=1, . . . , J
are suitably calculated as follows:
o.sub.j=.gamma..sub.j(net.sub.j, b.sub.j)
net.sub.j=.SIGMA..sub.iw.sub.ijx.sub.ir.sub.if.sub.ij
.gamma..sub.j=net.sub.jb.sub.j
o.sub.j=(.SIGMA..sub.iw.sub.ijx.sub.ir.sub.if.sub.ij)b.sub.j
[0128] If the user does not agree with the results of consultation,
he can modify the value of the effect in question, the modification
suitably being a graphical modification, by using the ACFM. He can
instruct the model that he believes the response from the model
ought to be higher or lower. Such a situation may arise if the
user's mental model has evolved or changed due to some reason, or
if he thinks that the aggregate outputs are not a correct
representation of his mental model. Before any changes are made to
the model the user suitably has the opportunity to assign a
confidence level, c.sub.j, to his new position. The system
preferably automatically adapts to the user's new position by
readjusting its internal parameters.
[0129] Turning now to FIG. 15, a connectionist representation of
the architecture for computing the j.sup.th effect is provided. For
the entire s-model with I causes and J effect, J nets will suitably
exist in parallel. When a user wants to modify the model he
supplies the desired value d.sub.j corresponding to the j.sup.th
effect. The difference D.sub.j between the desired value and the
computed value o.sub.j forms the basis for changes to the judgment
values .DELTA.w.sub.ij of the model. The user preferably has an
option of specifying his confidence c.sub.j in the proposed
modification. The relative importance r.sub.i of the I causes is
also considered. The change in judgment .DELTA.w.sub.ij is
preferably computed as follows:
.DELTA.w.sub.ij=D.sub.jr.sub.ic.sub.j, i=1, . . . , I
.DELTA.w.sub.ij=(d.sub.j-o.sub.j)r.sub.ic.sub.j.
[0130] The new judgment value at time t+1 is calculated as
follows:
w.sub.ij(t+1)=w.sub.ij(t)+.DELTA.w.sub.ij
[0131] Where:
[0132] w.sub.ij(t+1) is the new judgment based value for the
relationship and
[0133] w.sub.ij(t) is the current judgment based value for the
relationship.
[0134] Turning now to FIG. 16, a simplified connectionist schematic
illustration of the model of FIG. 15 is provided.
[0135] Once the s-model is built, the model and data contained
therein is manipulable and viewable just as is the e-model. That
is, all the functions described in the design of the e-model, such
as optimization, relative weightings, and graphing, etc., are also
available in the s-model.
[0136] FIG. 17 is a screen showing consultation results and faceted
decision hyper-plane surface of s-model
[0137] FIG. 18 screen showing two different manipulable graphical
views of the relationships and the results of corner solutions
optimization
[0138] FIG. 19 is a screen showing an interactive graphical display
of the relationships and the adaptive conjoint feedback mechanism
for cybernetic control or teaching the system graphically by
pushing and pulling on the graph.
[0139] FIG. 20 is a screen showing cascaded windows where each
window shows relationship vectors for all causes in one effect's
space.
[0140] FIG. 21 illustrates the overall design of the s-model
according to a presently preferred embodiment.
[0141] FIG. 22 is a high level diagram illustrating some of the
functionality of the s-model.
[0142] The present invention is suitably configured for use on a
stand-alone system, or in a network environment, whether
peer-to-peer, client-server, or N-tier, and is suitably
configurable to accommodate a plurality of users, all of whom
contribute to either a single mental model or a plurality of mental
models. In this manner, the present invention is suitably capable
of describing a mental model on an organizational level. In other
words, the present invention is suitably configurable to create an
organizational memory.
[0143] Turning now to FIG. 23, there is provided a system diagram
of the present invention in a network environment. The system
comprises a data transport network 2300 illustrative of a LAN or
WAN environment in which a preferred embodiment is provided, such
as a packet-switched TCP/IP-based global communication network. The
network 2300 is suitably any network and is preferably comprised of
physical layers and transport layers, as illustrated by a myriad of
conventional data transport mechanisms like Ethernet,
Token-Ring.TM., 802.11(b), or other wire-based or wireless data
communication mechanisms as will be apparent to one of ordinary
skill in the art.
[0144] Placed in data communication with data transport system 2300
is a Server 2302 which is suitably any Server for accommodating
selective query support, selective data access, data archiving, and
the like, as will be appreciated to one of ordinary skill in the
art. Preferably, the Server 2302 is a database server. One or more
Clients, such as representative Clients 2314, are also placed, or
selectively placed, in data communication with the data transport
system 2300. The Client 2314 is preferably configured to interact
with Server 2302 as will be appreciated by one who is skilled in
the art. It should be noted that the Clients 2314 are suitably
thick clients or thin clients, additional servers, PDA's, or any
equipment capable of interacting with Server 2302 to send and
receive data. Thus, a data path between Server 2302 and the one or
more clients, such as representative Client 2314, are in shared
data communication.
[0145] The Server 2302 preferably comprises a network interface
2308 through which the Server interfaces with data transport system
2300. The Server 2302 also preferably comprises internal storage,
which is suitably in the form of RAM and is suitably embodied as a
hard disk, optical storage, removable memory, or any other
preferably rewritable storage as will be appreciated by those
skilled in the art. Preferably stored on internal storage 2310 are
an E-model 2304, and S-model 2312, and a graphics server 2306. The
graphics server 2306 is suitably any graphics server software as
will be appreciated by those skilled in the art. The presently
preferred embodiment utilizes the Chart FX software package. The
E-model 2304 and S-model 2312 are suitably embodied as computer
readable code on a computer readable medium. In the presently
preferred embodiment, the E-model 2304 and S-model 2312 are
compiled C++ software components. It will be appreciated by those
skilled in the art the E-model 2304 and S-model 2312 are suitably
any computer readable code for supporting the functionality
described herein.
[0146] FIG. 24 illustrates the structure of the E-model 2304 and
the S-model 2312 according to the presently preferred
embodiment.
[0147] Although the preferred embodiment has been described in
detail, it should be understood that various changes,
substitutions, and alterations can be made therein without
departing from the spirit and scope of the invention as defined by
the appended claims. It will be appreciated that various changes in
the details, materials and arrangements of parts, which have been
herein described and illustrated in order to explain the nature of
the invention, may be made by those skilled in the area within the
principle and scope of the invention as will be expressed in the
appended claims.
* * * * *