FOBISS_Intel_whitepaper

Artificial Intelligence

for business processes’ optimization

WHITE PAPER

FOBISS BV

Beukenlaan 60

5651 CD Eindhoven

Netherlands

[email protected]

Introduction

Artificial intelligence (AI) concerns the study and development

of intelligent machines and software. The focal problems include

the development of software that can reason, gather

knowledge, find patterns, generalize, discover, associate, plan

intelligently, learn, communicate, adapt to new situations.

AI allows users of data to automate and enhance complex

descriptive, predictive and prescriptive analytical tasks that,

when performed by humans, would be extremely labour

intensive and time consuming. Thus, AI can have a significant

impact on the role data plays in deciding how we work and how

we conduct business.

Companies as well as public bodies seek to enhance their

competitive advantage by better understanding ever-growing

amounts of data. AI offers the technology and methodology to

do so, which is why the market for AI-based tools and

applications is growing rapidly.

Trends

Artificial intelligence can find use in many different sectors. In

any industry that deals with large amounts of data, techniques

and technologies based on artificial intelligence can be of value.

AI solutions are by nature multidisciplinary, involving computer

science, mathematics, statistics and philosophical thinking. It is

the basis for software that supports, facilitates and improves

analytics and decision making.

In the last years a significant progress has been achieved by

solving complex business processes’ analysis, optimization and

control problems through the using of various methods of

artificial intelligence. These methods include artificial neural

networks, fuzzy systems, multi-agent systems, evolutionary

computing, swarm intelligence and the virtual plant paradigm.

While individual models of biological and natural intelligence

have been applied successfully to solve real-world problems, the

current trend in artificial intelligence is to develop hybrids of

already known paradigms combining various intelligence

AI is the basis for software

that supports, facilitates

and improves analytics

and decision making.

techniques, since no one paradigm is superior to the others in

all situations. In doing so, we can use respective strengths of the

chosen algorithms or the models and eliminate weakness of

some other individual components.

Solving the Optimization Problem

Unfortunately, because of relatively complicated tools,

intelligent systems have only limited application examples in

improving of real business processes today, even though they

have been already very successfully used for various engineering

applications.

Our proposed general application schema of artificial

intelligence algorithms and methods is an attempt to

bring the intelligent systems’ methodology, techniques

and tools to the business community. Solving practical

tasks of business processes’ analysis, optimization and

control.

The following general schema includes all the innovative

methods and algorithms of artificial intelligence, which are

necessary for business processes optimization.

Figure 1. General application schema of AI methods and

algorithms for business processes optimization

The data and information flows from real or simulated processes

are collected using advanced multi-agent systems. This

information is used to create predictive models of the process.

Predictive models are used to find optimal decisions and actions

for business’ process optimization. Because of various process

constraints and complicated process models, advanced

optimization tools are involved here to find the best decisions

and problem oriented solutions.

“Virtual Business” tools included in the general schema allow the

users to test various process optimization and analysis

algorithms and “IF… THEN” scenarios, before they are making

costly and risky experiments with actual business processes.

For on-line monitoring and analysis of business process data, as

well as for detection of possible faults and unexpected behavior

of the process general application schema includes special tools

for intelligent data analysis. Because of high computation load it

is suggested for the companies to realize most computation

tasks of business optimization using cloud computing approach.

Cloud computing is a technology that uses Internet and powerful

remote servers to obtain business processes optimization

solution. This technology allows for much more efficient and

user-friendly computing by centralizing storage, memory,

processing and bandwidth.

The most important parts of this general application schema

are:

Intelligent predictive models

Intelligent optimization tools

Intelligent analysis tools

1. Intelligent predictive models

Today’s business world is driven by customer demand.

Unfortunately, the patterns of demand vary considerably from

period to period. This is why it can be so challenging to develop

accurate forecasts. Forecasting is the process of estimating

future events, and it is fundamental to all aspects of

management.

The new generation of AI technologies help improve the

forecasting process including such applications as product

demand, employee turnover, cash flow, distribution

requirements, inventory, manpower forecasting etc.

The proposed general application schema involves an innovative

approach for building business process models. We called it

“Hybrid Predictive Modeling” technique and it is based on

combination of fundamental models, operator knowledge and

data driven models.

“Hybrid Predictive Modeling” technique

Business process optimization requires a lot of knowledge about

the business processes. The classical way of representing

business process knowledge in science and engineering is to

use mathematical models based on first principles. This requires

a thorough understanding of mechanisms dominating the

business process dynamics. However, many essential details of

the various business processes are not yet so well understood in

order to construct well working fundamental mathematical

models. Hence, to establish valuable models additional

resources must be exploited.

Every-day experience shows that a great deal of more or less

quantitative knowledge about the business processes is

available that, so far, could not be represented in a form of

fundamental mathematical models. Thus, it is straightforward to

look for possibilities to incorporate this knowledge into

alternative kind of numerically evaluable process models and

knowledge-based prediction models.

The basis of knowledge-based prediction models is knowledge

from business process experts. Therefore the proposed

application schema should include fuzzy expert system

technique to form easy the basic rules for formulating of

prediction models for the business process. Also, data from

already running business processes cover a wealth of hidden

information about business process dynamics.

In our application to develop a data driven models we use a

special category of artificial neural networks – flexible neural

networks. This approach allows choosing automatically the

structure of neural networks adaptable for complexity of the

problem to be solved. For processes with high-dimension

inputs variables we proposed an original modification of deep

belief networks to construct the predictive models with high

generalization properties. To predict simple time series

processes Caterpillar methodology of singular spectrum analysis

(SSA) models are used.

The proposed hybrid modeling approach allows exploiting

collected data in very efficient way by using various sorts of

artificial neural networks and procedures to build data driven

predictive models.

Experience showed that neither fundamental process models,

nor heuristic descriptions, or various artificial neural networks

alone are sufficient to describe real business processes

accurately enough to make efficient process optimization. In

order to meet this requirement, all available knowledge should

be activated. In particular the information hidden in the

extended measurement data records from the process under

consideration must be exploited. Hence, procedures are needed

to simultaneously capitalize the available mathematical

modeling knowledge: fundamental mathematical models, the

information hidden in process data records as well as the

qualitative knowledge gained by the process operators through

their experience. This can be achieved with the ´Hybrid Process

Modeling´ technique.

Figure 2. Elements of hybrid modelling approach for predictive

modeling tasks

Extended knowledge embraces more information about

business process and creates opportunities for the development

of more accurate predictive process models. For the reason that

the real business processes usually are non-stationary, the

parameters of the hybrid models should be adapted on-line,

when new observations from the real processes reach the

databases.

2. Intelligent optimization tools

The developed prediction models are the basis for business

processes’ optimization. For the reason that the predictive

models, objective functions and optimization constraints usually

are very complicated in real business processes it is difficult to

apply traditional optimization and search methods for process

optimization tasks. For such cases the proposed general

application schema should be equipped with new established

optimization methods.

The optimization methods are required to have possibility to

combine evolutionary programming methods, genetic

algorithms, ant colony optimization and swarm intelligence

methods. Many successful applications already known in

practice have demonstrated, that these advanced optimization

methods are very powerful new tools for the investigation of a

wide range of optimization problems in real business processes.

3. Intelligent analysis tools

The research and applications of AI techniques in data analysis

include all areas of data visualization, data pre-processing

(fusion, editing, transformation, filtering, and sampling) and

database mining techniques. The proposed general application

schema advocate to focus on methods that can assist in

selecting and extracting the best ‘‘features’’ or condition

indicators from the whole data pool, that contain as much

information as possible.

These indicators present a reduced version of the original data

and preserve characteristic features of the process. Now, the

created indicators can be used to detect “unexpected behavior”

of the processes and to identify process faults.

Various data mining algorithms for process analysis could be

implemented in a proposed general application schema. The

most useful of them are methods based on Principal

Component Analysis (PCA), nonlinear (kernel) PCA, Self-

Organizing Networks, Deep-belief networks and fuzzy logic.

Example of AI application in business

FOBISS CM - Cash Management solution

for the Banking industry

Effective cash management is a crucial success

factor for banks or outsourcing companies.

FOBISS CM software offers for banks a Cash

Management solution to automate and improve

cash supply chain efficiency: cash logistics, ATM

and branch cash management (forecasting,

planning and optimization).

Using FOBISS CM software banks can benefit by

having more prime retail space available in

branches, freeing up staff to focus on customer

facing activities, shorter cash lead times,

maximize ATM availability and minimize running

costs.

FOBISS CM system employs artificial intelligence

techniques to control and optimize entire cash

cycle for different types of cash-points devices.

FOBISS CM system allows performance of

intelligent cash-point monitoring, resource

forecasting, cash delivery

management/optimization and business

processes’ analysis.

Traditional mathematical models have some

limitations when applied in cash management

systems (non-linearity, non-stationarity,

unknown relationships.

Artificial intelligence methods (neural networks,

fuzzy logic, evolutionary computation, swarm

intelligence) give more possibilities to design

and implement advanced cash management

systems.

Application areas of FOBISS Cash

Management

● ATM networks

● Automated teller safes

● Cash deposit systems

● Cash recycling systems

● Bank branches networks

Most important blocks of FOBISS cash management software are:

Intelligent predictive models of the business processes: employed techniques - support vector

machines, artificial neural networks, fuzzy logic.

Intelligent optimization tools: employed technique – advanced search algorithms, simulated annealing

methods, genetic and evolutionary optimization techniques.

Intelligent communication tools: employed techniques – multi agent systems.

Intelligent analysis tools (on-line detection of “unexpected” process behavior, and detection of process

faults): employed techniques – PCA, kernel PCA, advanced clustering algorithms, Self-Organizing

Networks and Deep belief networks.

Figure 3. Overall structure of the FOBISS CM system

Figure 4. Overall structure of the FOBISS CM system

Figure 4. Functionalities of the FOBISS CM system vs. other CM systems

MONITORING

FOBISS CM system uses special multi-agent based technologies to

collect the necessary transaction data and to present the

transaction results in a user friendly graphical environment (cash

demand/supply plots, statistics);

Some special monitoring functions are realized in the FOBISS CM

system:

Ranking of all ATMs in network based on cash demand/supply

volume;

Real-time cash state in every ATM and alerts connected with

high, normal and low level of cash;

Statistics of cash demand/supply for every ATM (day, week,

month);

Real-time cash demand/supply plots together with statistical

averages;

Correlation plots (cash demand/supply) of neighbors’ ATMs.

FORECASTING

Exploratory data analysis techniques are used to choose the most

important input variables for forecasting algorithms. Historical data

records, seasonality, holidays, local events are analyzed by forming

input variable groups for cash demand/supply forecasting.

Realized Forecasting algorithms:

● Flexible adaptive artificial neural networks (ANN);

- For every ATM or branch unit an ANN is employed to learn

complex relationships between inputs variables and cash

demand/supply;

- Every new data record is used to retrain the network;

● Adaptive fuzzy expert systems (expert clone);

- Experts’ knowledge is involved to form the initial cash

demand/supply forecasting rules using user friendly fuzzy

logic interface;

- Initial system is then adapted by using historical data and

real-time observation

● Adaptive support vector regression;

- Support vector regression is a special modification of support

vector machines techniques dedicated for solving of

nonlinear regression problems;

- SVR technique is well suited for starting phase of Cash

Management system, when the amount of data records for

ANN training is limited;

- Given training data, the support vector regression solves a

regression curve construction problem, where input variables

are mapped to a higher dimensional linear space by the

special kernel function Φ

OPTIMIZATION

FOBISS CM system reduces the cost of cash holdings and

replenishment by estimating optimal cash loads and loading

intervals for each cash point.

Factors which can be taken into consideration by process

optimization:

● Market interest rates;

● Cash supply and Logistics costs;

● Insurance costs;

● Physical constraints (min, max cash amount);

● User defined constraints;

Optimization techniques implemented in FOBISS CM system:

● Empirical combinatorics,

● Simulated annealing algorithms,

● Evolutionary programming methods

Special case: optimal routing for cash upload

Implemented technique: swarm intelligence – ant colony

optimization. Modeling of pheromone depositing by ants in their

search for the shortest paths to food sources is employed for the

creation of advanced shortest path optimization algorithm.

INTELLIGENT ANALYSIS

Intelligent analysis and early detection of the unexpected

behavior of the ATMs (or other cash-point devices) is

important for efficient functioning of networks;

Because of high service costs, it is very expensive to employ

human operators to supervise all ATMs (>1000) in an ATM

network;

FOBISS CM system includes an automatic identification

procedure based on auto associative artificial neural

networks (AANN) to supervise continuously the ATM

networks;

Auto associative Neural networks are trained using

advanced deep-auto-encoding procedure (Restricted

Bolzman machine, RBM);

Training procedures induce the neural network to model

correlations in the input data to reproduce the input data at

the output with minimal distortion.

Compression of information by the bottleneck results in the

acquisition of the correlation model of the input data,

useful for performing of further data analysis.

Trained auto associative neural network maps the input

variables into the space of the nonlinear correlation model

and the squared prediction error (SPE) of this mapping can

be used to detect unexpected behavior in the inputs.

Bank and ATM network savings

using FOBISS CM

Case 1

Due to improved analytic, decision

making and planning capabilities, a

bank with 6000 ATM network, saves

11,5 Million Euro yearly.

Case 2

After FOBISS CM deployment the bank

(185 branches) managed to decrease

dormant cash level by 41%.

Case 3

Once FOBISS CM was deployed, the

bank (1100 ATMs) started planning

cash delivery more accurately.

Consequently, planned and emergency

visits decreased by 16% and 64%

respectively.

CONCLUSION

The potential of Artificial Intelligence for

organizations is enormous. The AI market is

continuing to grow and the way businesses

operate will very soon take up a whole new

meaning, resulting in fewer employees

required while significantly improving bottom

line results.

AI technologies allow companies to automate

and improve complex descriptive, predictive

and prescriptive analytical tasks. AI helps to

understand associations between different

information flowing through companies and

can suggest relevant information to the right

person at the right moment for timely

decision-making. AI applications are intended

to reduce costs, improve customer satisfaction

and productivity and increase revenues.

FOBISS employs various artificial intelligence

techniques for business process and resource

planning, optimization and control. The

efficiency of the proposed technologies was

proved in various pilot projects. The

implemented techniques are modular and can

be suited according to special demands of

users. Systems functionality is easily

configurable and new methods and techniques

are straightforward to be incorporated in the

existing system configuration.

www.fobiss.com

FOBISS is the developer of advanced analytics software applications

which allow to automate, improve and distribute proactive decision

making across organizations in dynamic demand-driven industries.

The solution employs unique artificial intelligence methods to

optimize operational processes, efficiently plan resources and manage

risks.

© FOBISS BV and affiliated entities.

No part of this publication may be reproduced, stored in a retrieval

system or transmitted in any form by any means, electronic,

mechanical, photocopying, recording or otherwise, without the prior

permission of the publisher, FOBISS.