You are on page 1of 5

Machine Learning for Expert Systems in Data

Analysis
Ezekiel T. Ogidan Kamil Dimililer Yoney Kirsal Ever
Department of Software Engineering Department of Electrical and Electronic Department of Software Engineering
Faculty of Engineering Engineering Faculty of Engineering
Near East University, Nicosia Faculty of Engineering Near East University, Nicosia
Mersin 10 Turkey, North Cyprus Near East University, Nicosia Mersin 10 Turkey, North Cyprus
Email: ezekieltijesunimi.ogidan@neu.edu.tr Mersin 10 Turkey, North Cyprus Email:yoneykirsal.ever@neu.edu.tr
Email: kamil.dimililer@neu.edu.tr

Abstract—Imagine a system that automates the job of data large amounts of data at a fast rate and helping to arrive at
science in such a way that if given sales data of a store nationwide a decision quickly. Also, taking into consideration the fact
can analyze the rate of sales of a product relative to its supply and that its knowledge pool can be populated with information
demand and other necessary variables and is able to decide on the
best places for the company to open new stores and help them to from many experts, it offers the services of these many
decide on the best possible ways to distribute their products. Data experts, as opposed to just one expert. In some cases, Expert
Science is an interdisciplinary field that is very much like data systems have even been used to predict future problems.
mining and Knowledge Discovery in Databases (KDD), involving This shows that Expert systems are a indeed a valuable tool.
the analysis of data to make useful inferences and deduction. This has resulted in Expert Systems being incorporated into
However, taking into consideration the information overload of
this age, it has become a pressing need to automate the process of various fields and practices ranging from warfare and space
analyzing this data. At the point of this need is where data science exploration to business, credit verification, capital budgeting
meets machine learning. Machine learning being a powerful tool analysis and even health care.
for automation can be merged with data science and analysis to In this paper, we would discuss the Expert system tech-
make for a more effective faster way to analyze data. In this nology and split its development methodologies into eleven
paper, an application of expert systems for data analysis would
be discussed as a case study. different categories and discuss how machine learning factors
in each of the methodologies. We would also be discussing an
I. I NTRODUCTION application of an expert system that uses artificial intelligence
to provide in-call behavioural guidance to agents and measure
An Expert System is a system that emulates the deci- customer perception in real time over the phone calls.
sion making ability of a human expert in a restricted do-
main(Giarratano & Riley, 2002). It can also be seen as an II. E XPERT S YSTEM T ECHNOLOGY
intelligent computer program that uses knowledge and infer- The technology of and Expert System consists primarily of
ence procedures to solve problems that are difficult enough two components: the knowledge base and the inference engine.
to require significant human expertise for their solutions. The In applications of Expert Systems however, a user interface is
system studies patterns and occurrences in the data and is required for interaction between the user and the system.
able to draw inferences (Feigenbaum, 1988). In some cases,
the system might even explain the logic behind the inference. A. Knowledge Base
In an article titled ”Putting Expert Systems to Work” Expert Systems are developed with the intention of making
in the March 1988 issue of the Harvard business review, the system have the same level of expertise as a trained
Expert Systems were described as programs that mimic the professional in the field for which the system is built. As
the thinking of human experts that would otherwise perform a result, the system needs a knowledge base, or database of
analysis of a situation or data. The System is equipped with all the information and facts that makes an expert abundantly
the functionality to capture large proportions of data to build knowledgeable in their fields. Part of the development process
a pool or database of knowledge. All of this data is what is for the engineer would include deciding on which methods
analyzed by the system and patterns are studied to have an would be used to acquire this knowledge and represent it in
understanding of the data and help in making decisions of the system.
inferences in later situations. An adequate knowledge base that would be needed for the
Expert Systems have since been giving a competitive edge system to run with high similarities to an expert would include
to organizations that have incorporated it into their operation both factual and heuristic knowledge. The factual knowledge
processes. It offers the advantage of allowing analysis of entails all information and fact that is generally accepted and
978-1-5386-4184-2/18/$31.00 2018
c IEEE agreed upon by the experts in that field, while the heuristic
knowledge entails more flexible or variable information that For example, IF the traffic light is red, THEN stop; IF the
is gotten either by experience of irregular or uncommon traffic light is green, THEN go. In this example, the system
situations, personal judgment or even guessing. now has knowledge of how a traffic light works. This structure
relates information on the ”IF” part - also known as the
B. Inference engine
premise or condition - and relates it to an action or inference
This is the part where the analysis happens. This involves on the ”THEN” part.
all mechanisms employed in generating the conclusion based A rule might have multiple conditions joined by keywords
on the knowledge stored in the system. There are a number of like AND (for conjunction) and OR (for disjunction) or both.
approaches or methodologies that are employed by engineers Operators such as is, are, is not, and are not can also be used
in the development of an inference engine for an Expert to assign a value to an object in the condition.
System. They would be discussed in the next section. Rules can be used to represent:
C. User Interface 1) Relations: IF temperature is greater or equal to 38.0 C;
THEN you have a fever
This simply refers to the mechanisms that the user uses to 2) Recommendations: IF you have a fever; THEN see a
interact with the system. This may also include the mecha- Doctor;
nisms used to modify or update the knowledge base of the 3) Directive: IF You go to see a Doctor; AND The Doctor
system. prescribes medicine for you; THEN take your medicine;
Groups of rules can also be put together to develop a
strategy.

Fig. 1. General Structure of an Expert System

Fig. 2. Structure of a Rule-based Expert System


The Explanation subsystem is used by the system to explain
the logic or reasoning behind the inference. This is an optional
In some cases, you might have more than one rule with
part to an Expert System.
the same IF part in the knowledge base. These rules with the
III. E XPERT S YSTEM M ETHODOLOGIES same IF part constitute what is know as a conflict set. In these
The need for proper representation of knowledge in Expert cases, it is up to the inference engine to determine which rule
Systems is to allow the inference engine to quickly and easily to execute. The method used by the inference engine to decide
run operations to analyze the information and give correct and which of the rules should be executed is known as conflict
understandable results. The differences in the methodologies resolution.
used creates a distinction between Expert Systems, we would In some of these cases, the inference engine simply executes
discuss this approaches now. the rule with the highest priority. This is based on the
assumption that the rules have been organized in a certain
A. Rule-based Expert Systems order of priority in the knowledge base. This usually works
This is the most used form of knowledge representation. well in Expert systems with around 100 rules.
In rule-based systems, the systems are designed to carry out In some other cases, the most specific rule is executed.
analysis by a series of complicated ”if...then” rules that allow This means that the rule that addresses a particular condition
them to solve difficult one-of-kind problems. The Knowledge specially is executed as opposed to one that addresses the
is represented in form of these rules. situation as part of a general group. This is because more
specific rules are expected to process more information than what can be likened to learning by experience. The process
general rules. of gathering data to be classified in the learning process is
Another approach would be to execute the rule in the knowl- known as feature extraction. It involves extracting key features
edge base that makes use of data that has more recently been or unique values from an entity to be processed or learned from
entered in the database. This approach would consequently in the system. It is usually followed by feature selection, which
require that dates of entry for each fact in the database be is where the necessary features for the application are selected.
represented. This is a very important aspect of machine learning because
While rule-based systems might be the most popular version a well designed and implemented classifier could be made to
of Expert Systems, they lack the ability to learn from experi- perform poorly if the features are not properly selected.
ence. They cannot automatically update their knowledge base In feature extract, the curse of dimensionality refers to
based on new information and they stick to the rules always. a a situation where the number of features or dimensions
taken into consideration for a classification are so much and
B. Artificial Neural Network Systems cause the network to take more time to train. Dr. Charles
This is a methodology of Expert Systems that dwells primar- Isbell of the School of Interactive Computing, Georgia Tech
ily on machine learning. It takes advantage of the mechanisms states that ”As the number of features or dimensions grows,
of neural networks to simulate the human brain and emulate the amount of data we need to generalize accurately grows
its functions. In this paper, we would focus on Expert Systems exponentially”. The increase in dimensions in which data is to
built with Neural Networks. be represented in causes an exponential increase in the amount
In neural networks, nodes are developed to take input in the of data represented. As more dimensions are added, you get
same way that dendrites in biological neurons receive electro- to a point where the value or accuracy offered by adding these
chemical stimulus. The output of these nodes is also likened new dimensions is no longer worth the computational and time
to the signal output of the axon in a biological neuron. cost of training the system.

Fig. 3. Biological Neural Network Unit (dendrite)

Fig. 5. The exponential increase in data for each new dimension or feature
added

C. Fuzzy Expert Systems


Fig. 4. Artificial Neural Network Unit (node) showing similarities to its Fuzzy logic refers to a form of logic of half truths, where
biological counterpart truth values can range between 0 and 1 so as to handle
situations where the value is neither completely true nor
Neural Networks learn when data is passed through them. completely false. In this approach, the system is allowed to
They take note of patterns and and learn these patterns in be flexible and to give answers that are not exact. This is
born out of the ideology that real life problems are not always Within the fuzzy inference process, we have to first select
straightforward with simple true or false answers. It essentially propositions in the conditions part, or IF part of the IF-THEN
employs creative decision making processes. rules. We then apply truth values to the extracted rules. For
each of these rules, we also assign a level of importance to
the conditions to define the extent to which their outcomes
affect the final decision. These could be values such as Very
Important or Moderately Important.
Some other forms of Expert Systems classified by their
methodologies or approaches include Object Oriented method-
ology, Case-based reasoning, Modeling, System Architecture,
Intelligent agents, Ontology and Database methodology.
IV. E XPERT S YSTEMS WITH N EURAL N ETWORKS
One of the major applications of Neural Networks has been
Fig. 6. Basic Structure of a Fuzzy Expert System
pattern recognition. The network is able to learn patterns in
large amounts of data and associate this data to an inference
In fuzzy Expert Systems, like in neural networks, the or conclusion. This is of course similar to the general problem
necessary information for the knowledge base is extracted solved by Expert Systems, where the Expert System studies
from the training data set and the knowledge base is developed data and creates a knowledge base with which it makes
and represented in the form IF-THEN rules. The knowledge inferences of decisions. The neural network approach also has
base in this case would be in the form of a fuzzy neural the added advantage of allowing the knowledge base to be
network. A fuzzy neural network is simply a system that adjusted and updated based on comparison between the output
uses the estimation or approximation functionality of a neural inferred and the expected output
network to obtain the parameters of a fuzzy system. In an application where data on population growth is en-
tered, the system would be able to observe the birth and death
rates over a specific time and make a conclusion on whether
there is a population increase or decrease. With a set of IF-
THEN rules, the system would also be able to suggest ways to
stabilize the population growth rate based on the conclusions
derived.
In the development of a Expert Systems using neural
networks, we need to first identify the problem and define
it properly, so we can identify the necessary attributes or
features needed to solve the problem and avoid the curse
of dimensionality. Next, the values for these features are
Fig. 7. A Fuzzy Neural System Showing the Extraction of Fuzzy IF-THEN identified or obtained either by interviews, surveys or already
Rules existing databases. This is what forms our training data set.
Possible outputs would also be identified. the neural network
Fuzzy Inference Process refers to all the processes used to is trained using this data set, and the network consequently
map the input of the system to the output which is a final develops a knowledge base.
decision made based on the mapping or inference. Below is a
2 input, 3 rules, and 1 output rule system used to determine V. C OGITO AS A CASE STUDY
how much should be given as a tip to a waiter. Cogito is a system developed by cogito corporation that
performs in-call voice analysis and delivers real-time guidance
to agents and insight to managers. It leverages the functionality
of Artificial Neural Networks for behavioural analytics.
The system is based on research by behavioural analysts
between 1999-2007 at the Human Dynamics lab at MIT. The
researchers were able to prove that in human communication,
there are social signals passed between the people communi-
cating. They were also able to prove that these signals could
be detected by computers and used to draw inferences on
human behaviour. In 2007, the Cogito corporation was created
to bring this findings to the market.
From 2007 to 2011, Cogito created an artificial intelligence
Fig. 8. A Fuzzy Rule System platform and ran tests with behavioural models to interpret
human communication and detect psychological states. This VII. C ONCLUSION
research was funded by DARPA (Defense Advanced Research There are many methodologies that have been used in the
Projects Agency) and a number of other corporations. development of Expert Systems for over three decades now.
From 2012 to 2014, Cogito was deployed within health care While all having their advantages, the approach using neural
management programs at health and insurance companies and networks has proven to be the most effective and efficient way
thousands of interactions were analyzed, generating millions of developing durable and reliable Expert Systems that would
of data points. this helped to increase the effectiveness of the not become obsolete over time.
behavioural models used with Cogito. In this paper, we have discussed Cogito, a neural network
Since 2015, Cogito has been able to give in-call behavioral Expert System as a case study and shown how neural networks
guidance to agents and a real-time measure of customer per- can improve the work that is already done by Expert Systems.
ception for every phone conversation. In the time it has been Employing this methodology in Expert Systems for many
operational, Cogito corporation has reported 10% reduction in fields would further improve efficiency in those fields.
handle time, 16% increase in first call resolution, 14% increase
in customer satisfaction, 90% increase in customer feedback R EFERENCES
and 12% increase in deal close rates. [1] Liao, S.H. (2005), Expert System Methodologies and Applications – a
Cogito’s development process adhers to the development Decade Review from 1995 to 2004 Expert Systems with Applications,
28(1), pp 93-103.
process highlighted in the previous section. These steps are: [2] Poeschl, S., Lieb, J.,Wirth, F., & Bauernhansl, T., (2017), Expert Sys-
tems in Special Machinery: Increasing the Productivity of Processes in
A. Identify the Attributes/features Commissioning Procedia CIRP, 63, pp 545-550.
[3] Giarratano, J. G., (2002) , Expert Systems Principles and Programming
In the case of Cogito, the attributes that were Identified USA: PWS publishing.
and considered necessary for the system to make accurate [4] Feigenbaum, E., McCorduck, P., & Nii, H. P., (1988), The Rise of the
Expert Company New York: Times Books.
conclusions were: [5] Engin, G., Aksoyer, B., Avdagic, M., Bozanl, D., Hanay, U., Maden, D.,
• Energy & Ertek, G., (2014), Rule-based Expert Systems for Supporting University
Students Procedia Computer Science, 31, pp 22-31.
• Interruption
[6] Wiig, K. M.,(1994), Knowledge management, the central management
• Empathy focus for intelligent-acting organization/ Arlingtong: Schema Press.
• Participation [7] Mitra, R. S.,& Basu, A., (1997), Knowledge representation in MICKEY:
An expert system for designing microprocessor-based systems IEEE
• Tone
Transactions on Systems, Man, and Cybernetics.
• Pace [8] Dhaliwal, J. S.,& Benbasat, I., (1996), The use and effects of knowledge-
based system explanations: theoretical foundations and a framework for
B. Identify Values for the Attributes/features empirical evaluation Information Systems Research, 7, pp 342362.
[9] Choua, S. M., Leeb, T. S., Shaoc, Y. E., & Chenb, I. F., (2004), Mining the
In Cogito, these values were derived during the research and Breast Cancer Pattern Using Artificial Neural Networks and Multivariate
development period as part of the behavioural models. More adaptive regression splines Expert Systems with Applications, 27, pp
133142.
values were also gotten when it was introduced into health [10] Karabatak, M., & Ince, M. C., (2009), An Expert System for Detection of
and insurance companies Breast Cancer Based on Association Rules and Neural Network Expert
Systems with Applications, 36, pp 34653469.
[11] Bishop, C. M., (1996), Neural Networks for Pattern Recognition Ox-
C. Identify the Outputs ford: Clarendon Press.
The expected possible outputs were also defined. some [12] Haykin, S., (1994), Neural Networks, a Comprehensive Foundation
New York: Macmillan College Publishing Company Inc.
outputs were, agitation, confusion, stress, among others. [13] Jamshidi, M. A., Titli, A., Zadeh, L., & Boverie, S. (1997).Applications
With these processes, the network was trained and a knowl- of fuzzy logic: Towards high machine intelligent quotient systems Upper
edge base was formed with which the system now observes Saddle River, NJ: Prentice Hall.
[14] Hayashi, Y.,& Imura, A., (1990), A Neural Expert System with Auto-
conversations and is able to give the opinion and advice of an mated Extraction of Fuzzy If-Then Rules and Its Application to Medical
expert behavioural analyst. Diagnosis Proceedings: First International Symposium on Uncertainty
Modeling and Analysis, pp. 489-494.
VI. D ISCUSSION [15] Kottai, R. M., & Bahill, A. T., (1989).Expert Systems Made with Neural
Networks International Journal of Neural Networks, 1(4), pp 211-226
Cogito serves as a modern practical example of how Ex-
pert systems that make accurate conclusions effectively can
be developed using neural networks for machine learning.
Another added advantage to using neural networks for Expert
Systems is the fact that it further develops its knowledge base
as it works, unlike in rule-based expert systems, fuzzy expert
systems, among others. In fact, much like a human expert,
a neural network Expert System learns more with practice. It
becomes more of an expert as it works. It adjusts its knowledge
base based on the differences between its output and the
expected output.

You might also like