THE EVOLUTION OF INTELLIGENT SYSTEMS IN THE

MINING INDUSTRY
John A. Meech

The Centre for Environmental Research in Minerals, Metals, and Materials,

The University of British Columbia, Department of Mining Engineering,
Vancouver, British Columbia, CANADA
E-mail address: cerm3dir@mining.ubc.ca

Abstract

This paper reviews the evolution of computer systems based on the so-called "intelligent"
technologies now being applied in different application areas throughout the mining, minerals, and
materials industries. These systems have emerged from the field of "Artificial Intelligence" in
which expert systems, fuzzy logic, artificial neural networks, genetic algorithms, and agent-based
software have dominated. The Mining industry has been particularly receptive to these methods
since so many of our operations and processes are understood and controlled in empirical ways
that lend themselves well to the use of intelligent technologies. In addition there are few industries
with the myriad of heuristics evident in mining; for example:
- Nature does not make uniform orebodies nor ones that can be modeled simply;
- Unit operations tend to be batch or semi-continuous which are more difficult to
model and control;
- The traditional approach to problem-solving is empirical and such experiential
knowledge can be captured directly into Intelligent Systems.

INTRODUCTION

As the attributes of personal computing hardware (speed, memory, storage capacity,

resolution) have doubled every 18 months or so since the 1980s, our society has reached
a point where no serious performance limitations exist for "intelligent methods" and the
computational complexities are now embedded within or subsumed beneath the
Human-Machine Interface. As a result, these approaches can be applied to study and
solve extremely complex and intricate problems beyond the ability of the human mind to
handle in a time frame appropriate for process control. Process control has traditionally
tried to maintain a system at a set-point for as much time as possible in response to upsets
or disturbances in load variables. Nowadays, the set-points themselves have become
disturbances with updates occurring at increasing frequencies as communication and
measurement cycles have sped up to bandwidths previously unimaginable.

Data Management is a major issue today in complex process control. Can we take
advantage of so much data using filters or sensor-fusion techniques to increase the
performance of systems that may consist of loops at levels higher in the control hierarchy

presented at
Mineral Process Modelling, Simulation and Control Conference
Laurentian University, Sudbury, Ontario
June 5-7, 2006
 2

than ever before? Techniques are needed to deal with "long" delay times in local loops or
ones caused by the transfer of information to remote locations that previously provided
supervisory control at time frames measured in minutes to hours, but today operate in
seconds to minutes. The future of Intelligent Systems will integrate hardware and
software with the field of robotics playing a dominant role. As machines take over more
of the "routine" thinking, this provides time now to address tasks that are important, but
of lower priority. Complexity Analysis will become the next generation of control
systems with software seeking new relationships within the large amounts of data being
generated. Automated assembly of hardware parts to create self-replicating systems will
become important over the next decade as systems possess the ability to "self-heal" using
redundant sensors, actuators, and computer hardware.

A BRIEF HISTORY OF FUZZY LOGIC

It can be argued that Fuzzy Sets have been in use in Mining Engineering and its related
disciplines since the very beginning of the industry thousands of years ago. One can find
references in De Re Metallica in 1556 (one of the first books ever printed) on the use of
linguistic terminology to characterize mining variables [Agricola, 1556], but the formal
application of Fuzzy Logic to reveal the mathematics behind this linguistic terminology
did not occur until the mid-1980s with the seminal paper by Harris and Meech [1987]
that described a fuzzy approach to the control of crushing plants. Since that time, use of
Fuzzy Control and Fuzzy Set theory has expanded rapidly into virtually all areas of the
industry including geology, mining, metallurgy, and control of environmental pollution.

Mining is one of the oldest professions. Since early woman and man began using stones
to crush food and to throw rocks to chase off predators or kill prey for food, people have
been mining rocks and minerals for all kinds of use. Stones were crafted into weapons
and tools until it was discovered that when placed in fire under the right conditions,
components of the rock could be extracted to produce metals. This led our ancestors into
the Iron Age and the Bronze Age. As Mankind improved the ability to communicate
using speech sounds eventually developing words to share thoughts with friends and foes,
new methods evolved to extract rock and ore more easily to create more effective
products for agriculture, hunting, and protection. As a science, mathematics came along
much later, so it is reasonable to assume that language (Fuzzy Logic) as a method to pass
on ideas predates any application of formal mathematics.

The roots of mathematics lie in ancient Egypt and Babylonia, spreading rapidly into
ancient Greece where it was translated into Arabic and enriched by computational ideas
from the Indian sub-continent. Later, the science passed to the Romans and entered
Western Europe where within a relatively short period (200-300 years), methods of
computation spread around the world. While not wishing to disparage the importance of
mathematics in providing rapid ways to compute solutions with precision and accuracy,
many of the in-grained methods of mining, smelting, providing heat and shelter, and
producing new products, were well-established prior to widespread use of mathematics.
 3

By far, the most significant development in mathematics was provision of a firm

foundation in logics [Fauvel and Gray, 1987] by the ancient Greeks in the centuries just
prior to Euclid (~300 B.C.). These foundations gave mathematics more than a certainty–
they provided tools to investigate the unknown [Euclid, 295 B.C.]. Despite the pre-
existence of "Fuzzy Logic", what was lacking was the ability to make useful and correct
decisions in the absence of knowledge or data. The science of Greek logic was bivalent in
form–things were either true or false, although "the paradoxes" were identified even in
the beginning as problematic for bivalent logic. Despite this deficiency, the world
continued to evolve as the workers–farmers, fishers, and miners continued advancing
their work habits with linguistics central to their method of communication and to the
passing on of ideas (old or new). As mathematics entered the field, formalisms
(especially geometry and trigonometry) were applied to characterize relationships among
variables, but retention of the old empirical models has prevailed even into modern times.

Perhaps because of the empirical "culture of mining", Fuzzy Set Theory was able to
proliferate quickly once a formal scientific basis was given by Zadeh in 1965.
Application of this theory established a rationale for relative weights of importance used
to characterize underground rock masses, to select underground mining methods, to
design excavations from knowledge about the orebody shape and extent together with the
physical conditions of the surrounding rock masses for many years. The first successful
application of "Fuzzy Sets" occurred in the field of process control with Mamdani's
famous paper on the control of a laboratory-scale steam engine in 1975. Shortly
thereafter, the Danish cement industry applied the technique developed by Mamdani to
control a cement kiln. In 1976, Blue Circle Cement and SIRA in Denmark developed a
cement kiln controller that represents the first documented industrial application of Fuzzy
Logic Control [Yen, 1999]. Blue Circle was later taken over by the F.L. Smidth Group
who then grew this first industrial application into a company called FLS Automation
(interestingly enough, FLS can stand for either F.L. Smidth or for Fuzzy Logic Systems).
Virtually all cement kilns in the world today use a FL-based control system.

FL has allowed development of successful methods to move processes and operating

practices closer to the point of instability and/or failure. This has meant a closer approach
to "optimal" solutions. As the environment changes, control (or operating practice) can
adapt to maintain desired targets for longer time periods. FL has a self-adapting property
to mimic directly how a person "thinks" about the problem-space.

With clinker production, process lag times are measured in tens of hours resulting in poor
response of manual control especially when operating personnel are inexperienced or
preoccupied with other duties. Often, the term "experience" refers to recognizing
conditions that lead to failure. As such, "experience" actually derives from having
"failed" and not wanting to be in that condition ever again. With the long lag times that
characterize processes such as cement kilns, the ability of an operator to interpret
instruments that predict a future "failure" in 10 to 20 hours requires considerable skill that
may not be acquired from a single "bad" event. The complexities that lead to problems
can be entangled in ways that require more time and effort than are available.
 4

On the other hand, a good Mining Engineer designs a mine to fail–unlike buildings or
bridges that must survive for tens or hundreds of years. The stand-up time of a mine
opening is supposed to range from a few hours to as long as several months depending on
the rock properties and the scheduling of mining activities. To keep the cavity open for
longer periods of time is generally uneconomic since support or expensive maintenance
work is necessary. This has been the condition of mining since time began, although over
the past generation, the importance of worker health and safety has become paramount
(with the exception of coal mining in the People's Republic of China where ~4,500
people are reported to die in mines every year). In North America, Europe, and Australia,
mining is now a safer activity than that of construction. It has been a hard-fought battle to
achieve such statistical improvements and there is still considerable room for additional
advances [Hall, 1990], but health and safety are supreme concerns of mining activities in
many developed countries replacing production as the first priority of work.

HOW FUZZY LOGIC THEORY ENTERED THE MINING INDUSTRY

The seminal paper in mining described the application of Fuzzy Logic to the control of a
secondary crushing plant [Harris and Meech, 1987]. Although crushing is a continuous
process, these plants are subject to a large number of discrete upsets ranging from alarms
warning of the presence of metal or wood in the ore, to planned shutdowns for daily
maintenance. "Mother Nature" was rarely kind when she created orebodies–variations in
hardness and feed size can be considerable necessitating close attention to how each
individual crusher is performing. A circuit failure can be very expensive and instruments
available to monitor these upsets are relatively crude and susceptible to fouling by dust,
mud, and other ore contaminants. This is precisely the type of plant in which FL excels,
i.e., one that is subject to complex heuristic upsets, one that has non-linear relationships
among its key variables, and one in which sensor technologies are lacking.
Current Draw Current Draw Current Draw Current Draw Current Draw
HIGH MEDIUM HIGH OK MEDIUM LOW LOW

OR AND AND AND

Screen Fine Screen Fine Screen Fine Screen
Bin Level Bin Level Bin Level Bin Level
HIGH YES NO YES NO YES NO
OR
Chamber
Level
HIGH

NEGATIVE NEGATIVE NO POSITIVE POSITIVE

BIG SMALL CHANGE SMALL BIG

FEED RATE CHANGE =

WEIGHTED AVERAGE OF OUTPUT FUZZY SET SUPREMUMS
BASED ON CORRESPONDING DEGREE OF BELIEF

Figure 1 – Secondary Crusher Fuzzy Logic Control System [Harris and Meech, 1987].

The initial work [Harris and Meech, 1987] demonstrated how FL outperforms manual
operation and achieves better results than unsupervised PI control. Set-point adjustment
 5

of PI control provided gains that matched those from Fuzzy Control, and the potential to
fine-tune the system to achieve even higher performance was considered a real
possibility. Later work [Meech and Jordan, 1993] established the importance of
adaptation with respect to hardness and feed particle size. In addition to achieving a 25%
increase in plant throughput, the use of a Fuzzy Confidence Level to delete rules with low
Degrees of Belief was shown to influence system stability positively. Figure 1 shows the
simplicity of the system to control a secondary crusher (one stage of the entire process).
By allowing Fuzzy definitions to change in real-time, significant improvement occurred.

A group at the University of Alabama [Karr, 1991, Karr et al., 1990, Karr and Gentry,
1992a/b, Karr, 1993] began applying soft-computing techniques in the late 1980s–early
1990s. Their initial work used FL to characterize the appearance of flotation tailing
material to give plant operators a more consistent analysis. Later, this work was extended
to pH control (a particularly difficult non-linear problem), and then to control of a
flotation process. They used a genetic algorithm (GA) to alter the shape of Fuzzy Sets
used in a given rule base, especially for pH control, and so began the evolution of hybrid
systems that employ neural network technologies as well as GA.

MINERAL PROCESSING PLANT CONTROL

Cámara [1999] has reviewed Fuzzy Systems in Mining that provide tools to improve and
optimize productivity of processing and metallurgical plants as well as manage
maintenance requirements. He pointed to the success of these systems to provide a safer
working environment as well as the flexibility to change a process "on the fly". One
system operating across adjacent plants providing cross-over control in a coordinated
fashion to meet complex corporate objectives. Stability and adaptability of both plants
was significantly improved as process supervision attempted to reach a higher goal.

Savolainen [1998] has reported on kiln control using FL. The work compared Fuzzy with
multivariable control in terms of: reduced GHG emissions; reduced energy consumption;
increased refractory brick life; an easier, more stable kiln operation; and development of
tools for remote operation. Fuel consumption was reduced significantly and temperature
peaks harmful to refractory life were eliminated. The burnt lime quality defined by
residual carbonate improved (lower amount with less variation). See Figure 2.

Raatikainen [1998] demonstrated an advanced control system for cement plants and
limestone quarries that helped improve operations by saving raw materials and improving
control using a distributed XRF-analyser with a Material Management System based on
FL. The system was developed with sub-suppliers, such as General Electric R&D, who
played a key role in designing the FL controller.

PCE Engineering in Finland [Kauhanen and Mattila, 1998] developed a Fuzzy System to
keep temperature and moisture stable during blending and curing of concrete under
changing weather conditions to produce concrete products of uniform quality. The system
 6

calculates heating times and estimates temperature and moisture for aggregate materials.
The amount of cold and hot water to achieve target temperature and moisture content is
then determined. In a related cooperative R&D effort, Peltonen [1999] created a simple
MatLab toolkit to configure Fuzzy controllers for various automation tasks where poor
performance was obtained with conventional methods–such as control loops with
nonlinear processes. Fuzzy systems can be built using either expert knowledge or process
data. The toolkit has shown promising results in pilot trials at two Finnish paper mills.

Figure 2 – Significantly decreased carbonate level and variance using

Fuzzy Logic in a Finnish lime kiln [Savolainen, 1998].

Applications have been built to control column flotation cells using FL in Japan
[Hirajima et al., 1996], Portugal [Carvalho and Durão, 1999], and Canada [Kosick and
Harris, 1988]. This latter reference is the seminal paper on applying ES for real-time
flotation control. Work in Sweden has applied FL to operate a process to remove
phosphorous from magnetite [Su, 1998]. FL has also been applied for hydrocyclone
control to prepare ore for flotation [Wong et al, 2004]. Chilean researchers have applied
FL to control flotation circuits – both simulation models and real plants [A. Cipriano,
1999]. Baiden and Meech [1987] reported on using simulation to study aspects of the
mine-mill interface that cause production bottlenecks–the need for operator training and
good inter-department communication with respect to scheduled maintenance was
demonstrated using such models.

Several researchers have used FL to assess data trends in on-stream assays. Using a
windowing technique, trends were measured over various time horizons to provide input
into decision-making with respect to changing reagent addition [Poirier and Meech,
1993], [Kivikunnas, 1999].

While et al. [2004] reported that over 15 intelligent crusher control systems have been
installed in the mineral industry since the mid-1990 by Minnovex Technologies. These
systems used evolutionary algorithms to design and operate crushing plant circuits.
 7

Chilviet and Meech [1996] have reported on "qualitative modeling" to build intelligent
monitoring and control systems for SAG circuits.

At the Carajas Mine in Brazil, FL was used to control a tailing thickener [Santos et al.,
1995]. Savings in fresh water pumping and reduced addition of flocculant paid for the
system in two months. Other dewatering applications involve the modeling of a rotary
dryer by Finnish researchers that follows on from earlier work on lime kiln control
[Yliniemi et al., 2003].

With so much reliance on dependable instrumentation, the ability to predict failed sensor
readings from other data has been important. FL plays a role in performing these sensor
fusion functions [Mahajan, 2001]. Cifuentes et al. in 1995 developed an on-line
qualitative model of a semi-autonomous grinding circuit based on FL for use by mill
personnel to monitor and evaluate factors responsible for delays and production losses by
interpreting combinations of sensor signal trend patterns.

In Australia, BEC Engineering and FLS Automation installed a Fuzzy Logic-based mill
control system on the Mt. Rawdon Gold Mine SAG-mill circuit in 2002 to increase
circuit efficiency and improve throughput. The process consisted of a SAG mill-screen
operation with oversize mill discharge diverted to a pebble-crusher before returning back
to the SAG mill. The main throughput limitation was mill power draw and so, the system
goal was to maintain SAG Mill power as close to maximum as possible. The level of
instrumentation available was sparse although standard PID controllers were available to
adjust mill feed rate, water addition and sump level.

Figure 3 – Equigold's Mt. Rawdon SAG Mill circuit control loops.

The rule-based expert system uses FL to operate the circuit under normal conditions as
well as to recover control following an emergency shut-down or upset. Each sub-control
loop (power, tonnage, and sump-level) can be turned on and off (i.e., over-ridden) by an
operator. When active, the system cascades set-points to individual PID controllers. The
 8

system was well-accepted by all operators and gave a documented throughput increase of
3.5% (~6-month payback). The graph in Figure 4 provides an example of system
performance. Initially, changes in mill feed rate done manually caused large excursions in
power from the desired set-point. When the control system is turned on at 08:15, many
small alterations are made to the mill feed rate which reduced variations in mill power. It
is claimed that the FL-based control system is like having the best operator running the
mill 24 hours a day [F.L. Smidth Group, 2003] (see Figure 5).

Figure 4 – SAG mill performance at Mt. Rawdon Gold with

and without Fuzzy Logic-based advanced control.

Figure 5 – FLS Automation's ProcessExpert system for cement kiln operation.

(note the use of soft sensors to predict "difficult-to-measure" parameters
such as Blaine (specific surface area) and free lime determinations.)
- from F.L. Smidth Group, 2003

FL is not limited to process control, but is also applied to train new operators [Meech,
1990, Ikonen and Najim, 1997], trouble-shoot control loops [Chang and Chang, 2003],
and design gold recovery plants based on mineralogical analysis [Torres et al., 2000].
Monitoring systems are also important applications [National NEMO Network].
 9

GEOLOGICAL AND GEOGRAPHICAL APPLICATIONS

The major applications in geological and geographical systems relate to the generation of
maps. These include mapping landslides [Zhu et al., 2004], soil characteristics, [Zhu et
al., 2001], uncertainty in GIS [Zhu, 2005], [Carranza and Hale, 2001], [Zeng et al., 1997],
[Waters and Evans, 2003]. An excellent comparison of the use of crisp classification
versus Fuzzy classification is given by Shalan et al [2003]. They point to the advantage of
using Fuzzy boundaries and discuss some pitfalls in using FL without understanding
Fuzzy Set definitions.

Considerable literature exists on the characterization of mineralized zones, for example,

gold in the Philippines [Carranza and Hale, 2001], copper in Iran [Ranjbar et al., 2002],
and copper/gold in Mauritania [Eden et al., 2002a/b]. Evolutionary mapping of satellite
images of a national park in Sardinia using Fuzzy techniques is monitoring the extent of
development [Manca and Pireddu, 2003]. Fuzzy modeling has also been applied to
interpret geophysical data [Bardossy and Duckstein, 1995], while another group created a
Fuzzy Expert System to teach mineral identification [Nagel and Meech, 1995].

MINING APPLICATIONS

Figure 6 – Knowledge representation in the MMS Expert System [Clayton et al., 2002]

Built on top of the UBC Mining Method Selection algorithm, MMS [Clayton et al., 2002]
is a knowledge-based system that incorporates FL in its analysis. The MMS System
modifies the UBC approach by considering uncertainty associated with the boundaries
between input parameter categories. Through a series of logical operations in the
knowledge base, an estimated Degree of Belief in each of the ranked mining methods is
calculated. The system was validated against two other selection programs and shown
able to provide additional advice to a design engineer by provision of a Degree of Belief.
Figure 6 shows the Fuzzy representation that characterizes parameters used to select an
 10

appropriate method. The use of words to characterize each variable has been used by
expert mining engineers well-prior to the formal application of FL.

The Helsinki University of Technology [Koivo, 1999] in partnership with Sandvik-

Tamrock, has built an Artificial Neural Network with an FL-based system to determine
the bucket payload of a front-end loader to provide a mine with immediate information
on material flow from the mine. This can result in improved control of the whole mine
production process and allow monitoring of individual tasks and machines to calculate
salary bonuses and pre-plan maintenance. Weighing during movement is not very
accurate and requires data pre-processing to account for system nonlinearities. During the
project, innovations to solve this weighing problem were discovered leading to a patent
application in which a Fuzzy-Neural system is the key.

Rock Mechanics is an important area of modern mining in which a mine is monitored for
rock failures on a continuous basis with stope design using empirical methods based on
past experience. Stress in rock bolts is measured regularly with data being interpreted
using FL [King, 1996]. Other rock mechanics applications using FL methodologies have
been reported in India [P. Singh, 1998], [T. Singh et al., 2005].

Telerobotics in underground mining [Tuokko et al., 1994, Sulkanen and Tuokko, 1994] is
now being applied in many parts of the world with a trend toward fully-autonomous
operation not far-off. A number of applications are using FL to control vehicle operation
[Hemami, 1994a], to interpret obstacle detection data [Polotski, 1995], to conduct path
planning and tracking [Cohen et al., 1998], [Hemami and Polotski, 1998], and to optimize
mucking and bucket loading [Hemami et al., 1994a/b], [Hemami, 1994b/c].

In open pit mining, excavation systems [Bissé et al., 1995], [Babienko, 2001], [Wang and
Lever, 1994], [Hemami, 1995a/b] have been developed using FL with emphasis on
cooperative or agent-based hardware/software [Niemelä, 1994], [Lacroix et al., 1999].
FL-based expert systems are helping to select open pit mining equipment [Basçetin,
2003] and mobile underground mining equipment [Papavasileiou et al., 2002]. An FL-
based production model [Huang and Kumar, 1993] is used at one mine to follow trends
and maintain steady operations. In longwall coal mining, an FL system was created to
control the load and speed of a coal shearing machine [Heyduk, 2001] allowing the
operator to remain in a remote position. Fuzzy control of a ventilation system has been
studied with great success. Such systems can direct air where needed and block-off areas
not requiring ventilation leading to significant cost savings and enhanced worker health
[Poanta and Dojcsar, 2001]. Coal blending and ash monitoring using FL have also been
developed in Polish coal mines [Cierpisz and Heyduk, 2001, 2002], Bydon, 2003].

ENVIRONMENTAL APPLICATIONS

Protection of the environment from mining activity has evolved over the past generation
from a small external movement looking for evidence to an essential department of every
 11

major company in the world. To receive a permit to mine today, one must plan for
closure before starting production. The plan must include all funding arrangements and
prove that the land will be restored following the operating cycle which may last for 10-
20 years. Monitoring the future environment is essential to ensure problems do not arise
after the company has left the region. This may include operating water treatment plants
to prevent heavy metals and acid waters from fouling nearby water courses. Fuzzy Logic
has an important role to play in creating automated systems. They provide an economic
incentive for a company to fulfill these obligations.

Metsä-Serla Oy, Kyro Board Mill and Neles Automation [Puhakka, 1996] developed a
fuzzy controller for an active sludge water treatment plant. The results were a reduction
in sludge age from 12 to 7 days leading to an increase in sludge removal during the bio-
sludge cycle. At the same time, the buffer capacity of the biosludge thickener increased
because of improved sludge-drying. With a 35% decrease in phosphoric acid use,
suspended solids decreased by 33% and the phosphorus load by 50%.

Virtually all mines attempt to recycle water as much as possible to avoid misuse of a
precious resource, particularly in arid or semi-arid climates. Wastewater management and
treatment is essential at all mine properties [Bongards, 1999], [Shrestha et al., 1996],
[Duckstein et al., 1994]. Prediction of ground water flows through waste dumps or into
open pits is being done using Fuzzy Systems [Scott, 1998] to attempt to avoid permanent
damage to surface soils [Komac and Sajn, 2001].

Prediction of pollution is an "art-form" fraught with considerable uncertainty as the

conditions that generate pollution can take years or decades to show-up since they depend
on the microbiology of waste piles and tailings dams. Many systems exist to perform risk
assessment [Ghomshei and Meech, 2000], [Veiga and Meech, 1995] with the most
successful ones based on Fuzzy Logic [Veiga and Meech, 1995b, 1997]. A fuzzy model
has been proposed to show how a new idea can migrate from a small group to become the
central theme of a society–this transformation is called Technological Evolution [Meech
and Veiga, 1998]. By monitoring the progress of such change, policy-makers can decide
on strategies to either promote or head-off the proliferation of a particular activity.

The mining industry is also characterized by small-scale mines. Over 100 million people
worldwide are either directly involved or indirectly reliant on such artisanal work [Veiga
and Meech, 1995a]. These activities are generally carried out in a disorganized fashion
with little respect for the environment leading to severe damage and pollution,
particularly in gold mining where mercury is used [Meech and Veiga, 1998]. A fuzzy
expert system called HgEx was developed for use by a variety of skilled personnel who
are working with these artisanal miners to attempt to improve conditions and reduce their
impact on society. Figure 7 shows how the system can deal with either measured data or
linguistic concepts to characterize observations at a mine site. Figure 8 shows how fuzzy
sets are interpreted from pH measurements causing rules to fire that can determine the
degree of danger in a particular environment subject to mercury emissions.
 12

Figure 7 – Process by which HgEx uses FL to Figure 8 – Fuzzy Sets used to describe pH
conclude on the extent of Hg pollution conditions of soil in assessing the
at a specific small-scale mining site Hg pollution danger at a particular
[Meech and Veiga, 1998]. mine site [Veiga and Meech, 1995c].

Figure 9 contains a simplified diagram of the modules used in HgEx to estimate the
degree of danger to organisms subject to Hg emissions in the Amazon. Note that the
model can be adapted using an "alpha" factor to characterize the economic, socio-
political and technical issues in a region that affect how an expert would conclude about
danger. In this way, the overall system can be adapted to other situations. For example
the presence of one small miner in the Amazon is really a "minor" problem compared to
his presence on a major North American river such as the Colorado or Fraser.

Figure 9 – Modules used in HgEx to predict the potential risk to biota in a

particular environment and mining site [Meech and Veiga, 1997].

In North America, there would be a loud reaction to this situation, while in the Amazon
the additional contribution is small and priority should be given to larger-scale operations
or regions where mercury use is "pandemic".

Figure 10 shows how the value of "alpha" affects conclusions about the danger or
concern in a region. The "alpha" factor is determined from a extensive, detailed analysis
of the elements that affect mercury pollution in a particular region or country or time. The
rules that determine "alpha" are depicted in Figure 11. Note the non-linear nature of this
 13

relationship in which the central region of the map forms a broad plateau with changes
occurring as the system approaches the edges of the graph.

Figure 10 – Linguistic descriptions of the extent Figure 11 – The rules used to calculate the
of Hg emissions depend on the value value of "alpha" in HgEx.
of "alpha". [Meech and Veiga, 1997]. [Meech and Veiga, 1997].

Work has been done to extend Eh/pH diagrams to interpret species domination in a
particular aquatic system at varying "redox" and pH conditions. Convention Eh/pH
diagrams show crisp boundaries between each region. Figure 12 shows a plot of this
analysis for mercury speciation. Each graph reflects a single concentration of each
reacting ion species (in this case, Hg+2, Cl-, OH-, and S-2).

Figure 12 – Eh/pH diagram for the Figure 13 – Effect of ligand concentration

equilibrium of Hg species on the boundary between Hgo
in an aqueous environment and Hg(H-1L)- in an Eh-pH
[Meech and Veiga, 1997]. diagram [Veiga et al., 1995].

Unfortunately in the real-world, there are many other species that can affect the
concentration of each of these species. For example, the presence of organics can rapidly
change the level of a dangerous species from low to high. Fuzzy Logic can be applied to
these thermodynamic calculations to provide a range of conditions under which these
changes are understood. Figure 13 shows how the boundary between the "relatively
benign" Hgo and the much more dangerous oxidized complex ion Hg(H-1L)- changes as a
 14

function of the complex ligand (L) concentration. With chloride present, the change can
be even more dramatic as shown in Figure 14 where the situation appears to be stable and
then shifts in the presence of organic material found in many parts of the Amazon.

Figure 14 – How the degree of belief in a dangerous situation can change from no danger to
certain danger at conditions in many regions of the Amazon [Veiga et al., 1995].

GENETIC ALGORITHMS AND EVOLUTIONARY SYSTEMS

Genetic Algorithms (GAs) are adaptive evolutionary methods used to solve large,
complex search and optimization problems [Goldberg, 1989] based on the genetic
evolution of biological organisms. Over many generations, natural populations evolve
according to the principles of natural selection and "survival of the fittest". By mimicking
this process, a GA can evolve solutions to real-world problems. The method works with a
population of individuals, each representing a possible solution. Each individual is
assigned a "fitness score" according to how well the solution solves the problem.
Individuals with high fitness values are given an opportunity to reproduce by "breeding"
with other "fit" solutions. This produces new individuals as offspring that share some
features from each parent. The least-fit members are less likely to be selected for
reproduction, and so, they die off. Selecting the best individuals and mating them to
produce a set of offspring produces a new population of solutions. With a well-designed
GA, the population converges quickly to the optimal solution.

Genetic Algorithms differ from more traditional optimization procedures in four ways:
1. GAs work with a coding of the set of parameters, not the parameters themselves;
2. GAs search from a population of points, not a single point;
3. GAs use payoff information directly, not derivatives or auxiliary knowledge;
4. GAs use probabilistic rules, not deterministic ones.

The coding of the solution is the most important step in designing a useful GA. Each
position on the chromosome string represents one variable in the solution space. The
value assigned to each position represents the state of that variable. Binary
representations are often used however, human DNA can take on one of 4 possible
values. The more levels used, the more complex (and fuzzy) the system becomes. So
typically, the value of each element is represented by a 0 or a 1 (on or off).
 15

GAs work from this rich set of database points by simultaneously climbing many peaks
in parallel, so the probability of being trapped on a false global peak is lower than that of
methods that move from point to point. The mechanics of a GA involves copying strings
and swapping partial ones. The simplicity of operation and the attributes of the effect
(speed and accuracy) are the main attractions of GA.

The next important step is defining the "fitness function"–a mathematical expression
characterizing the importance of variables in the solution space. System constraints on
variables can be dealt with within the "fitness function" by creating penalties in the total
fitness value. Weighting of the different elements in the fitness function can also be done.
A fitness function must be a non-negative measure and so the Maximum or Minimum
functions can constrain solutions to a limit of 0 or 1. A selection probability, Pi, is
assigned to each individual based on its fitness Fi , so the fittest individuals have an
increased chance of selection:
N
Pi = Fi / ∑
i =1
Fi (where N = population size) 1.

A simple GA that yields good results in many practical problems uses three operators:
Reproduction, Crossover, and Mutation.

Reproduction is a process by which individual strings are selected according to their

fitness value f . The function f is a measure of profit, utility, or goodness that is to be
maximized or a cost that is to be minimized. Selecting a chromosome according to its
fitness value means strings with higher (or lower) values have a greater probability of
contributing offspring to the next generation.

Crossover proceeds in two steps. First, members of the selected strings in the mating
pool are chosen at random. Next, each pair of strings undergoes crossover: an integer
position k along the string is selected at random between 1 and the string length (L) less
one [1, L-1]. Swapping all characters between position k+1 and L inclusively creates two
new strings. For example, if there are two strings:

101^011
110^100

selected for reproduction with a crossover position of 3, then the offspring created will be

101^100
110^011

Mutation plays a secondary role in operating a GA. Mutation is necessary since the
reproduction and crossover operations occasionally lose useful genetic material. The
mutation operation randomly chooses a single individual, randomly selects one position
on its chromosome string and transposes it from 0 to 1 or vice-versa. Mutation restores
diversity, but does not provide a logical approach to optimization, nor can it prevent a
 16

reoccurrence. Its use is important where a local minima (or maxima) traps the algorithm
and a new population member is needed to trigger the crossover operator on to a better
result. The probability of mutation is typically set to 0.01 to 0.001. Too high a mutation
rate can create a high influx of new genetic material upsetting the crossover process. A
time-dependent change in probability has demonstrated success with the initial value set
high and then declining as a function of the generation [Baker, 1985], [Bramlette, 1991]

Crossover
Crossover
Fitness Function

Crossover Mutation
Mutation
Crossover
Mutation

0
0 Generation Number

Figure 15 – Idealized effect of mutation operations on convergence of a GA.

ARTIFICIAL NEURAL NETWORKS

The concept of using the actual structure of the human brain to configure a system in
which input information is connected to output decisions has existed since the end of the
2nd World War. Concepts of an artificial neuron has progressed through several stages
with its roots grounded in neurological work done in the early part of the 20th Century.
From studies on the structure of nerve tissue, neurons were shown to be physically
separated cells connected to one another. Based on this work, McCulloch and Pitts [1943]
claimed that neurons with binary inputs and a step-threshold activation function were
analogous to first order systems but their simple model did not use connection weights.

In 1949, Hebb revolutionized the perception of artificial neurons. He proposed that when
an axon of cell A is near enough to excite cell B and repeatedly fires, a metabolic change
takes place such that A's efficiency, as one of the cells linked to B, is increased. This has
become known as Hebb's Rule which implies that when two neurons fire together, their
connection strengthens–an operation fundamental for effective learning and memory.

The McCulloch-Pitts model had to be altered to allow the adjustment of the weight of
each input. Rosenblatt [1958], using the McCulloch-Pitts neuron and the findings of
Hebb, developed the first Perceptron Model of the neuron which is still widely accepted
today. A Perceptron learns by weighting its inputs. The model is shown in Figure 16.

Each input is weighted and summed at the node with the total passing through to
activation function resulting in an output between 0 and 1. The inputs do not have equal
weights and the Perceptron can "learn" these weights through continued stimulation with
 17

data. The original activation function was a step-function (or threshold). However, other
functions such as Sigmoid, Piecewise-Linear and Gaussian activation have been
applied. Figure 17 shows how the Sigmoid function can be changed to provide variations
all the way through to a step change by increasing the size of the scaling factor.

Xo
INPUTS
Wjo
X1
Wj1
Yj
Σ f
OUTPUT
WEIGHTS

Processing Element
Wjn
(Perceptron)

Xn

Figure 16 – The Typical Perceptron Model of Figure 17 – Effect of scaling factor on the shape
a neuron (after Rosenblatt, 1958). of the activation function.

Unfortunately, the Perceptron by itself was limited to solving only certain mathematical
relationships. But Rosenblatt was so devoted to his Perceptron that he made the ill-timed
declaration that an error correction procedure will always yield a solution in finite time.
With this assertion, Rosenblatt essentially challenged head-on the symbol manipulation
projects being performed by others, i.e., Expert Systems.

In 1969, Minsky and Papert published a famous (infamous?) monograph entitled

Perceptrons: An Introduction to Computational Geometry that proved the model could
only solve linearly separable functions. They stated that this research was doomed to
failure because of these limitations–an equally inopportune remark. As a result, little
research was funded in the field. Rosenblatt died in an unfortunate boating accident in
1969 shortly after publication of this monograph–there are rumors he committed suicide.
In 1982, Hopfield demonstrated from work on the neuronal structure of the common
garden slug that ANNs can solve non-separable problems by placing a hidden layer
between the input and output layers [Hopfield, 1982], [Hopfield et al, 1983].

Since those heady days, a huge proliferation in ANN methodologies has occurred. Albus
developed the "Cerebellar Model Articulion Controller" in 1975 and this eventually
evolved into Agent-based behavioral systems that are becoming vogue today in the field
of Robotics. Rumelhart and McCelland's group at Carnegie-Mellon developed the most
famous learning algorithm in ANN–Back Propagation–which uses a gradient-descent
technique to propagate error through a network to adjust the weights in an attempt to find
the global error minimum [Ananthraman and Garg, 1993], [Burgin, 1992], [Jones et al.,
1987]. More recently, Sutherland [2001] developed a Holographic Neural Network (H-
Net) in which thousands of data points can be "enfolded" onto one neuron (Figure 18).
The method involves polar coordinate regression analysis in which each data point is
characterized as a complex number with the angle representing its value and the vector
 18

being its degree of belief or measurement uncertainty (Figure 19). The claim is made that
this technique is three-orders of magnitude faster than conventional connectionist ANNs.

Figure 18 – Cortical Cell block diagram in H-Net. Figure 19 – Multiple Pathways Defining
a Complex Scalar in H-Net.

ANN technology is actually an extension of Fuzzy Logic. When a rulebase is designed

using Fuzzy Sets, the degrees of belief emanating from each rule are equivalent to the
strengths of signals flowing between neurons in an ANN. In the case of known expertise,
the weights of recognized important inputs, sub-goals, or final conclusions are set to full-
strength, while those of unimportant inputs (according to the expertise) are set to zero.
Now if a neural network is designed to test this expertise on real data, the algorithm may
discover that some of the "important" links should be diminished while some of the
"unimportant" links should be strengthened. In this way, memories that link input facts to
outputs become more distributed while the ability to explain the reasoning is reduced.

HYBRID SYSTEMS

Finding ways to combine these techniques has predominated recent research. The initial
days of AI considered neural networks and expert systems as unique, separate fields.
Neural networks distribute memories and connections throughout the structure of the
system using data analysis while expert systems are based directly on expertise of a
human being who is willing to express linguistic expressions and relationships of the
knowledge. The former systems are black-box models that cannot simply explain their
knowledge to the world, while expert systems by their very structure (facts, rules,
relationships) have an inherent ability to explain and justify their decision-making.

Expert Systems struggled through the 1970s to develop solutions to significantly-sized

problems and once the AI community embraced FL as a technique integral to the process
of synthesizing uncertainty, success was realized. FL provides ways in which the number
of rules can be contained and the "curse of dimensionality" avoided. The relationship
between system size (number of rules) and I/O variable size (number of facts) can be
reduced from an exponential function to one based on simple multiples. The tuning of
Fuzzy Sets has always been viewed as a limitation by those who don't understand the
methodology. One must remember that all systems based on AI produce approximate
 19

solutions [Gonzalez and Perez, 1996]. Applying FL requires belief in a "tolerance for
imprecision" and so, despite the fuzzy definitions being "in error", methods to adapt these
relationships can be implemented to respond to other knowledge. FL, ANN or GA
methods can provide this adaptability [Satyadas and Krishna-Kumar, 1996]. These
approaches generate a variety of combinations of the basic technologies:

Fuzzy Expert Systems: ES with adaptable FL

Fuzzy-Neural Systems: FL with ANN-based adaptation of the Fuzzy Sets
Neuro-Fuzzy Systems: ANN with FL manipulation of link weights
Fuzzy-GA Systems: GA with fuzzy values (non-binary) [Herrera Lozano, 1996]
Neuro-GA Systems: ANN with GA choosing the best network of many solutions
Genetic-Fuzzy Systems: GA-based adaptation of the Fuzzy Sets [Lee and Takagi, 1996],

INNOVATIVE AGENT-BASED SYSTEMS

Mining companies in the Third Millennium must transform themselves into intelligent,
learning organizations able to cope with globalization of information resources. The main
problem is not access to information, but the ability to "mine" data and transform it into
useful strategic resources. [Szczerbicki and Gomolka, 1999]. As systems increase in
complexity, decomposition is the usual way to structure a problem. These atomized
structures consist of autonomous subsystems, each deciding on the information it receives
and sends [Gunasekaran and Sarhadi, 1997]. In the real-world, autonomous subsystems
consist of groups of people and/or machines tied together by service relationships.

An effective architecture [Davis, 1999] has the following features:

• The user can specify high-level tasks decomposable into more detailed execution
tasks according to an established hierarchy or distribution network;
• The user can plan and control at different resolutions of time and level of detail;
• The system can decompose complex behaviours into manageable sub-functions;
• The system allows a function to be distributed across several intelligent controllers.

An example of such a structure is the RCS design architecture suggested by NASA/NIST

shown in Figure 20. [Albus and Quintero, 1990], [Lumia, 1994], [Moncton, 1997]. The
Reference Model Architecture is based on Real-time Control Systems (RCS) developed
by NIST and has been widely adopted as a standard by NASA in its tele-robotic missions
and has served as inspiration for DoD's JAUS (Joint Architecture for Unmanned
Vehicles). In order to achieve the global goal in RCS, the global task is decomposed
according to the time scale of different components of the mission. This generates a
hierarchical control layer architecture. The bandwidth, as well as the resolution of spatial
and temporal patterns is planned so it decreases from bottom to top by about an order of
magnitude at each layer. Each layer consists of elements (nodes) of intelligence that
independently process their relevant sensory data, make decisions and control their
actions. These elements sense their environment and update their World Model. Each has
 20

its own unique World Model representation. Agent-based architectures have been
included in the Reference Model more recently as building blocks.

Sensor World Task

Processing Modeling Decomposition Time Scale

maps S1 M1 E1 operational
days to years
scheduling
S2 M2 E2 hours
object lists task Multiple
actions
state User minutes
variables S3 M3 E3
path
objective planning Interface
functions S4 M4 E4 seconds
dynamic
operations
program
files S5 M5 E5 milliseconds
servo
model plan and control
detect and
evaluatio execute
integrate
n
Figure 20 – NASA/NIST Standard Reference Modeling Environment [after Moncton, 1997].

Intelligent control agents run asynchronously and in parallel. To ensure completion of the
final goal under time constraints, Agents are organized into a hierarchy in which the top
layer is responsible for supervising overall task completion in a timely fashion. This layer
is the Main Supervisor Agent and all other activities are controlled by this agent. Most
real-world robotic systems and process control applications are dynamic with variables
operating at different bandwidths, so this layering of the architecture by time-resolution
removes interference from Agent demands for system resources. Bottom layers deal with
high bandwidth activities including interfacing with humans who may be sharing the
same space. Managing complexity, change, and disturbances are key issues in these
systems. A distributed, agent-based structure is an alternative to a hierarchy. The
cooperation within an agent-based structure with evolutionary schedulers allows a system
to handle complexity, reactivity, disturbances, and optimality issues simultaneously.

An agent is an "encapsulated" software entity with its own identity, state, behavior,
thread of control, and ability to interact with other entities including people, other agents
and "legacy" systems. An agent, whether real or virtual, can act on itself and on other
agents. Its behavior is based on observations, knowledge, and interactions with other
agents in the system or process. An agent has several important abilities–to perceive at
least a partial representation of its environment, to communicate with other agents, to
produce child agents. This knowledge of its own objectives and unique autonomous
behavior are often characterized as selfishness. [Monostori and Kádár, 1999].

Holonic Systems (Figure 23) are a relatively new paradigm in manufacturing akin to
agent-based systems. They consist of autonomous, intelligent, flexible, distributed,
 21

cooperative agents or holons [Valckenaers et al., 1994]. The word holon derives from the
field of holography–a holon is a part of a whole. The essential difference between an
agent and a holon is that hardware (instruments and actuators) can be included as part of
a holon whereas agents refer only to software entities (although not exclusively). Three
basic holons exist–resource, product, and order holons [Van Brussel et al., 1998].
These elements use object-oriented concepts to perform their duties. The most promising
feature is the transition from hierarchical to heterarchical systems. An object-oriented
framework to develop and evaluate distributed agent systems provides a model to
represent a plant containing different types of agents. (Figure 24).

Value
Judgment
Perceived Command
Objects & Situation Plan Plan Task
Events Evaluation Results Evaluation (Goal)

Update Plan
Sensory World Behavior
Processing Modeling Generation
Predicted State
Input Knowledge Command
Observed Database Actions
Input (Subgoals)

Figure 21 – An Agent-based Real-time Control System node.

An agent contains different, functionally-independent subagents. Each agent incorporates

a communication subagent to send and receive messages using a network protocol. A
resource agent involves a supervisor subagent to control real-world activities. Agents
contain a registration mechanism by which they access system hardware resources. Each
agent has a local knowledge and database to store information about machine capacities,
time intervals for different work, the groups of interest, etc. Data about the agent is
accessed through a communication subagent. (Monostori et al, 1998).

Agent-based software was invented to facilitate interoperability. There has been much
interest and development in "middle"-ware to deal with software that is already written–
so-called legacy software to allow it to remain in productive use. An agent is motivated
by intention (goal-oriented) and is modulated by its attention (prioritizing is a function of
static information as well as continually-measured dynamic data). A control agent
encapsulates a behaviour decomposed into subtasks of a behaviour-based nature that
react to environmental changes or action outputs from a decision-procedure analysis. The
action can be a message sent to another agent to perform a certain action or receive data.
Each agent has at least one active thread. Behavioural agents are feedback controllers
designed to achieve specific tasks/goals.
 22

Holonic System
Material flow Object B
Information flow

Holonifier Holonifier
Information Holonifier
Holon AGV
machine System Assembly
Object A’ Robot
“I want Object A
Object C”
Object C
Start
Customer Transport
Holon

Figure 23 – Heterogeneous Holonic Manufacturing System consisting of real and soft holons.
(after Monostori and Kádár, 1999)

Incoming Sent
message message
Information flow
Material flow
Output
Input
Message
Message Box Box

Message processing

Registration Communication
mechanism agent Knowledge
base

Local Resource supervisor

Database agent

Material
Processing

Figure 24 – Structure of a resource agent. (after Monostori and Kadar, 1999)

Behaviours connect sensors to actuators and receive input from and send output to other
behaviours. When assembled into distributed representations, behavioural agents can
 23

look ahead at a time-scale useful to the rest of the system. Large-scale, cooperative
teams, comprising interacting agents, offer capabilities beyond conventional software. An
infrastructure with these features uses small pieces of reusable code to solve problems via
interactions with other elements, rather than duplicating functions in other modules.

ROBOTIC APPLICATIONS

Intelligent robotic systems are developed within an architecture that enables intelligence
to be an integral part of the robot. Behaviour-based systems can express such intelligence
based on emergent behaviours of a complex system. When behaviours are implemented
by control agents, they express intelligent, flexible, cost-effective, modular, safe,
dependable, robust and user-driven qualities. The agents communicate with a human-
interaction agent who acts on behalf of the human at the internal software control
bandwidth, i.e., the agent's bandwidth is similar to that of other software, so it can close
control loops at low frequency with a human and at high frequency with software.

An important aspect of the continuous operation of a robotic system is its ability to

respond swiftly and modify its actual physical behaviour as new pieces of data are
perceived. Developing a robotic agent-based-system requires event-based monitoring
capabilities and management tools. To this end an algorithmic paradigm must be adopted
to support such demands based on online and event-driven ideas. An algorithm designed
to solve a given problem needs input data and parameter values from the outside
environment. The algorithm is not a fixed calculation, but rather, it adapts to changing
circumstances and its parameters. These situations are in a state of flux with regular
updates at bandwidths sufficient to address task requirements under control of the agent.

The underlying design of safe, robust and dependable robot systems operating in a human
environment and co-operating with people is integral to all aspects of robot R&D from
architecture development to key component functionality. Safety takes place by having a
robot continuously survey its surroundings looking for danger. The robot must possess a
self-monitoring ability and be able to shut down or repair itself, even partially, in the
event of a self-perceived emergency. They must be equipped with "e-safe" capabilities
that enable a distant human to implement immediate rescue action [R. Granot, 2003].
Robot systems cooperate with humans through an intelligent interface agent, so a human
can send change-orders on-the-fly providing responses to unexpected events. Activities
being monitored by humans in a telerobotic, supervisory mode of operation enable
communication between robot and human from a distance.

Advanced integrated modular robotics and the modular design and modelling of new,
versatile "plug-and-play" systems are developed in open-source reference architectures
with standardized hardware and software building blocks. These enable Agent-based
subsystems within a hierarchical-layered architecture as previously discussed. Installing
agents within RCS and JAUS is a major robotic advancement. Reusable agent code is
available as building blocks for various behaviours by their nature and definition. The
 24

capability of tele-robotic and human supervisory systems enables introduction of robots

into human environments. It extends the world of information processing by perception
of the physical environment using advanced multi-sensors with transfer of the analyzed
data as information to either humans or virtual agents for action.

Tele-robotic systems to operate underground and off-road vehicles is now being used by
a number of mines around the world (Northpark and Olympic Dam mines in Australia
and Inco in Canada). The future of these applications lies with autonomous systems. The
successful completion in October 2005 of the DARPA Grand Challenge suggests that the
required hardware and software are available today to create successful applications.
Safety is the top issue with a system reliability of 99.999% being required (5 minutes
downtime in one year). To achieve this level, agent-based software must be self-healing
and contain significant redundancy with respect to software components, instrumentation,
and final control elements. A diagram of such a system is shown in Figure 25.

FINAL WORDS

With the increasing pressures of globalization and environmental standards, mining

companies must examine alternative strategies to deal with these complexities and remain
sustainable. The software tools currently being used within the manufacturing sector have
potential to provide solutions in integrating systems across our different departments.
Although individual departments may find it difficult to communicate effectively due to
political situations or personality clashes, the types of computer systems that use these
intelligent architectures and components can pass over such constraints.

The field of AI has evolved from Expert Systems and Fuzzy Logic into hybrid systems
that include Genetic Algorithms and Artificial Neural Networks. These tools have
become embedded today within an overall distributed, reactive architecture known as
Agent-based systems. The applications of the future, particularly those which come from
the field of robotics will be built using these methodologies.

These methods can collect and store massive amounts of real-time control data for
decision-making at various corporate levels from direct-unit control to supervisory and
long-range planning. Corporations will develop simulation models inside of which many
different behaviours at many different time scales and spatial horizons will interact.
Failure to adopt these approaches will result in companies failing to recognize in a timely
fashion the inevitable change in high commodity prices. These innovations will be
necessary to sustain an enterprise successfully into the future.
 25

Figure 25 – Example of the Self-Healing Ability of a Mobile Robotic Control System.

ACKNOWLEDGMENT

The author is indebted to the many colleagues and students with whom he has worked at
the University of British Columbia, Queen's University, and at different mine sites
around the world. The patience and advice extended to him by his many friends and
colleagues has been a major factor in continuing to perform research in this exciting field.
 26

REFERENCES
G. Agricola, 1556. De Re Metallica, translated by Lou Henry Hoover and Herbert Hoover, Dover Pub. Inc.,
1950 (reprint of 1912 edition), Mineola, NY, pp. 672.
J.S. Albus, 1975. A New Approach to Manipulator Control: The Cerebellar Model Articulion Controller
(CMAC). Trans. ASME, Sept., 220-227.
J.S. Albus, R. Quintero, 1990. Towards a Reference Model Architecture for Real-Time Intelligent Control
Systems (ARTICS). Proceedings of ISRAM'90, 243-250.
S. Ananthraman, D.P. Garg, 1993. Backpropagation vs CMAC : Neural Network Control of a SCARA Robot,
Dep't. Mech. Eng. & Materials Sci., Duke University, Durham, NC, USA.
A.G. Babienko, A.W. Pokshin, 2001. Application of methods of Fuzzy set theory to monitoring and control
of excavators, ICAMC-2001: 14th Intern. Conf. on Automation in Mining, Tampere.
G. Baiden, J.A. Meech, 1987. Simulating the Mine/Mill Interface. Intern. J. Surface Mining, 1(3), 191-198.
J. E. Baker, 1985. Adaptive selection methods for genetic algorithms, Proc. 1st Inter. Conf. on GA, 101-111.
A. Bardossy, L. Duckstein, 1995. Fuzzy rule-based Modeling in Geophysical, Economic, Biological and
Engineering Systems, CRC Press, Boca Raton, FL, pp. 256.
A. Basçetin, 2003. A Decision Support System for Optimal Equipment Selection in Open Pit Mining:
Analytical Hierarchy Process, Istanbul Üniv. Müh. Fak. Yerbilimleri Dergisi, C.16, S.2, SS, 1-11.
E. Bissé, A. Hemami and E. Boukas, 1995. A Comparison of the Required Energy in Loading for Four
Scooping Strategies, Proceedings the Third International Symposium on Mine Mechanization and
Automation, Golden, Colorado, June, pp. 8.
M. Bongards, 1999. Controlling Biological Wastewater Treatment Plants using Fuzzy Control and Neural
Networks. 142-150, Computational Intelligence, Theory and Applications, 6th Intern. Conf. - Fuzzy
Days, Dortmund, Lecture Notes in Computer Science 1625 Springer.
M.F. Bramlette, 1991. Initialization, mutation and selection methods in GAs for function optimization, Proc.
4th Inter. Conf. on GAs, 100-107.
G. Burgin, 1992. Using Cerebellar Arithmetic Computers, AI Expert, 7, 33-46 .
S. Bydon, 2003. Supervisory Fuzzy Controller for Linear Control System, Dep't. Process Control, University
of Mining and Metallurgy, Kraków, Archiwum Process Control Club, pp.6.
M. P. Cámara, 1999. Fuzzy Logic System for Integrated Automation of Process Control and Management of
Total Production, Universidad Tecnológica del Perú, Lima. (translated)
http://www.monografias.com/trabajos16/sistema-Fuzzy-mineria /sistema-Fuzzy-mineria.shtml
J.M. Carranza and M. Hale, 2001. Geologically-constrained Fuzzy Mapping of Gold Mineralization
Potential, Baguio, Philippines, Conf. of Intern. Assoc. for Math. Geol. IAMG2001, Cancún.
T. Carvalho, F. Durão, 1999. Control of a flotation column using Fuzzy Logic inference, Fuzzy Sets and
Systems, 125(1), 121-133.
S.Y. Chang, C.T. Chang, 2003. A Fuzzy-logic based fault diagnosis strategy for process control loops.
Chemical Engineering Science, 58(15).
E.I. Chilviet, J.A. Meech, 1996. Intelligent systems to assist in semiautogenous grinding circuit supervision.
in Mine Simulation, Eds. G. Panagiotou, J. Sturgul, Proc. MineSim '96, 1st International Symp. on Mine
Simulation via the InterNet, Dec. 2-13, CyberSpace, Balkema, Rotterdam, p.39 + 14 html pages.
E. Cifuentes, J.A. Meech, A.L. Mular, H. Raabe and J. Mitchell, 1995. Qualitative Modeling of a SAG
Circuit at HVC, IFSA-95, 6th IFSA World Congress, Sao Paulo, Brazil, 477-480.
S. Cierpisz, A. Heyduk, 2001. Coal blending process control with Fuzzy controllers, ICAMC-2001: 14th
Intern. Conf. on Automation in Mining, Tampere.
S. Cierpisz, A. Heyduk, 2002. A Simulation Study of Coal Blending Control Using a Fuzzy Logic Ash
Monitor, Control Engineering Practice, IFAC, Mar.
A. Cipriano, 1999. Assessment of expert Fuzzy controllers for flotation plants. Minerals Eng., 12, 1327-1338.
C. Clayton, R. Pakalnis, J. Meech, 2002. Knowledge-based System to Select a Mining Method, in: Intelligence
in a Materials World. Eds. J. Meech, M. Veiga, Y. Kawazoe, S. LeClair, CRC Press, NY, 161-178.
P. Cohen, V. Polotski, M. Piotte, F. Melamed, 1998. Obstacle Detection System for Underground Mining
Vehicles, CIM Bulletin, 91, 75-79.
W.J. Davis, 1999. Distributed Intelligent Control of Complex Systems. Proc. IPMM'99 - 2nd International
Conference on Intelligent Processing and Manufacturing of Materials, Honolulu, HI, 1, 583-590.
L. Duckstein, A. Bardossy, A. Tecle, I. Bogardi, 1994. Fuzzy composite programming with application to
 27

wastewater management under changing physical conditions, In: Fuzzy Optimization, Recent Advances,
Eds. M. Delgado, J. Kacprzyk, J. Verdegay, M. Vila, Physica-Verlag, Heidelberg, 199-219.
K. Eden, F.M. Meyer, 2002a. GIS-based mapping in Mauritania. Mining Journal, 338(8690), 458-459 and
339(8700), 148-149.
K. Eden, F.M. Meyer, 2002b. Copper-Gold Potential Mapping in Mauritania. African Mining, 7(6); Brooks
Patrick Publications Pty Ltd.; pp. 18-25.
Euclid, 295 B.C. Euclid's Elements, Book 1: Fundamentals of geometry - Book 13: Regular solids:
http://aleph0.clarku.edu /~djoyce/java/elements/toc.html
J. Fauvel and J. Gray, Eds., 1987. The History of Mathematics, Macmillan, 1987: from D.E. Joyce, 1994.
http://aleph0.clarku.edu /~djoyce/mathhist /mathhist.html
S. Forouzi, J.A. Meech, 1999. Development of an adaptive artificial neural network for Brunswick Mining’s
copper/lead/zinc flotation circuit, Proceedings of IPMM'99, 2nd International Conference on Intelligent
Processing and Manufacturing of Materials, July, Honolulu, Hawaii, pp.8.
M.M. Ghomshei and J.A. Meech, 2000. Application of Fuzzy Logic in Environmental Risk Assessment:
Some Thoughts on Fuzzy Sets; Cybernetics and Systems; 31(3), 317-332.
D. E. Goldberg, 1989. Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley.
A. Gonzalez, R. Perez, 1996. Genetic Algorithms and Soft Computing: A Learning System of Fuzzy Control
Rules Based on Genetic Algorithms, Physica-Verlag, 202-225.
R. Granot, 2003. An agent based Human Robot Interaction in a supervised autonomous system, SPIE
AeroSense, Orlando, FL, paper #5083-45
A. Gunasekaran, M. Sarhadi, 1997. Planning and management issues in enterprise integration, Concurrent
Engineering: Research and Applications, 235-247.
A. Hall, 1990. Production Politics and the Construction of Consent: A Case Study of Health and Safety in
Mining, University of British Columbia, Dep't Sociology, Ph.D. Thesis, (J. Wayne - Supervisor).
C.A. Harris and J. Meech, 1987. Fuzzy Logic: A Potential Control Technique for Mineral Processing, CIM
Bulletin, 80(905), 51-59: originally presented at the 87th AGM of CIM, Vancouver, B.C., April 1985,
Paper No. 85-47, pp. 40.
D.O. Hebb, 1949. The Organization of Behavior. New York: Wiley, 62.
A. Hemami, 1994a. An Overview of the Elements of Automated Excavation, Proceedings of the World
Automation Congress - WAC'94, 1, Aug.
A. Hemami, 1994b. Motion Trajectory Study in the Scooping Operation of an LHD-Loader, IEEE Trans. on
Control Applications, 30(5), 1333-1338.
A. Hemami, 1994c. Study of Forces in the Scooping Operation of a Mechanical Loader, Trans. Canadian
Society of Mechanical Engineers, 18(3), 191-205.
A. Hemami, M. Mehrabi, R. Cheng, 1994a. Optimal Kinematic Path Tracking Control of Mobile Robot with
Front Wheel Steering, Robotica, 12, 563-568.
A. Hemami, S. Goulet, M. Aubertin, 1994b. Resistance of Particulate Media to Excavation: Application to
Bucket Loading, Int. Journal of Surface Mining and Reclamation, 8, 125-129.
A. Hemami, 1995a. A Fundamental Analysis of Robotic Excavation, J. Aerospace Engineering, 8(4), 175-179.
A. Hemami, 1995b. Comparison of a Logical Control and Fuzzy Logic Control for a Two-input Two-output
Mechanical System, Proc. Robotics and Knowledge-based Systems Workshop, Montréal.
A. Hemami, V. Polotski, 1998. Path Tracking Control Problem Formulation of an LHD-Loader, J. of Robotic
Research, 17(2), 193-99.
F. Herrera, M. Lozano, 1996. GAs and Soft Computing: Adaptation of GA Parameters based on Fuzzy Logic
Controllers, Physica-Verlag. 95-125.
A. Heyduk, 2001. Fuzzy Logic load and speed controller for a longwall shearer with inverter-fed haulage
drive system, ICAMC-2001: 14th Intern. Conf. on Automation in Mining, Tampere.
T. Hirajima, K. Oshima, M. Tsunekawa, Y. Otsuka, S. Kubo, and T. Imai, 1996. Fuzzy Control of Column
Flotation at Toyoha Mines, J. Min. and Mat. Proc., Inst. Japan (Shigen-to-Sozai), 112, 161-66.
J.J. Hopfield, 1982. Neural Networks and Physical Systems with Emergent Collective Computational
Abilities. Proc. National Academy of Sciences, USA, Vol. 79, 2554-2558.
J.J. Hopfield, D.I. Feinstein, R.G. Palmers, 1983. Unlearning has a Stabilizing Effect in Collective Memories.
Nature, 304, 158-159.
Y. Huang, U. Kumar, 1993. Application of Fuzzy Logic to estimate the working condition of a mine
production system: a case study, in: Proc. 23rd APCOM Conf., Montreal, Nov., 201-207.
 28

E. Ikonen, K. Najim, 1997. Use of learning automata in distributed Fuzzy Logic processor training. IEE Proc.
- Control Theory and Applications, 144(3), 255-262.
ISO10303-22, 1994. Industrial automation systems and integration: Product data representation and
exchange; Part 22: Standard data access interface specification.
W.P. Jones, J. Hoskins, 1987. Back-Propagation : A Generalized Delta Learning Rule. BYTE, Oct. 1987.
A. Karcher, J. Wirtz, 1999. PDM-based Virtual Enterprises - Bridging the Semantic Gap. IPMM'99 -2nd
International Conf. on Intelligent Processing and Manufacturing of Materials, Honolulu, HI, 1, 591-596.
C.L. Karr, 1991. Genetic Algorithms for Fuzzy Controllers, AI Expert, 6(2), 26-33.
C.L. Karr, D.L. Meredith, and D.A. Stanley, 1990. Fuzzy process control with a genetic algorithm. In: Eds.
R. Rajamani and J. Herbst, Control'90 - Mineral & Metal. Proc., SME/AIME, Littleton, CO, 53-60.
C.L. Karr and E.J. Gentry, 1992a. A genetics-based adaptive pH Fuzzy Logic controller. International Fuzzy
Systems and Intelligent Control Conf. (IFSICC'92), Louisville, KY, 255-264.
C.L. Karr and E.J. Gentry, 1992b. Real time pH control using Fuzzy Logic and genetic algorithms, Proc.
Ann. Meeting Soc. for Mining, Metallurgy, and Exploration, Phoenix, AZ.
C.L. Karr and E.J. Gentry, 1993. Application of Fuzzy control techniques to a chaotic system, Proc. Symp. on
Emerging Computer Techniques for the Min. Ind., SME/AIME, Littleton, CO, 371-376.
C.L. Karr, 1993. Real time process control with Fuzzy Logic and genetic algorithms, in: Emerging Computer
Techniques for the Mining Industry, Eds. B. Schneider, D. Stanley, SME/AIME, Littleton, CO, 31-38.
M. Kauhanen and T. Mattila, 1998. Intelligent dosing control system, PCE Eng. and Tampere Univ., Finland.
S. Khatami, J. Breadner and J.A. Meech, 2001. Unit Commitment for BC-Hydro's MICA Dam Generating
Plant using a Genetic Algorithm Approach, Intelligence in a Materials World, Ed. J. Meech, M. Viega,
Y. Kawazoe, S. LeClair, CRC Press, Boca Raton, FL, 295-302.
R.H. King, 1996. Fusing Data from Roof Bolter Sensors, Colorado School of Mines-USBM.
S. Kivikunnas, 1999. Overview of Process Trend Analysis Methods and Applications, European Network for
Fuzzy Logic and Uncertainty Modeling in Information Technology, Aug. 99.
H. Koivo, 1999. Neuro-Fuzzy application for on-line weight determination of a moving loader, Helsinki
University of Technology, Control Engineering Lab., Helsinki.
M. Komac and R. Sajn, 2001. Polluted or Nonpolluted –A Fuzzy Approach Determining Soil Pollution, Ann.
Conf. Intern. Assoc. for Mathematical Geology, IAMG2001, Cancún, Mexico, Sept. 6-12.
G. Kosick, C. Harris, 1988. Column Flotation Control at the Polaris Mine, Canadian Mineral Processors,
Ottawa, ON.
J. Koskinen, L. Yliniemi, K. Leiviskä, 1998. Fuzzy Modelling of a Pilot Plant Rotary Dryer. Proc. UKACC
Intern. Conf. on CONTROL´98. pp.4.
P. Lacroix, V. Polotski, P. Cohen, 1999. Decentralized Control of Cooperative Multi-Robot Systems,
Integrated Computer Aided Engineering, Wiley Interscience.
M. Lee, H. Takagi, 1996. GAs and Soft Computing: Hybrid Genetic-Fuzzy Systems for Intelligent System
Design, Physica-Verlag, 226-250.
R. Lumia, 1994. Using NASREM for Real-Time Sensory Interactive Robot Control. Robotica, 12, 127-135.
A. Mahajan, K. Wang, P.K. Ray, 2001. Multisensor Integration and Fusion using a Fuzzy Inference System,
IEEE/ASME Transactions on Mechatronics, 6(2).
E.H. Mamdani and S. Assilina, 1975. An experiment in linguistic synthesis with a fuzzy logic controller,
Inter. J. of Man-Machine Studies, 7(1), 1-13.
G. Manca, L. Pireddu, 2003. The Geopark in Sardinia: a Fuzzy spatial analysis for assessing the development
tendency, Proc. 7th Inter. Conf. on GeoComputation, Univ. Southampton, U.K.
J.A. Meech, 1990. Expert Systems for Teaching and Training; Mineral Engineering, 3(1), 15-30.
J.A. Meech and L.A. Jordon, 1993. A Self-Tuning Fuzzy Logic Controller, Minerals Eng., 6(2), 119-131.
J. Meech, M. Veiga, 1998. An Adaptive Fuzzy Model to Evaluate Technological Evolution. Minerals Eng.,
11(7), 597-604.
J. Meech and M. Veiga, 1997. Predicting the Impact of Mercury Pollution with a Fuzzy Expert System. Proc.
97-IEEE SMC, Orlando, Florida, Oct., 2, 1056-1061.
M. Minsky, S. Papert, 1969. Perceptrons, MIT Press, Cambridge, MA, USA.
S.P. Monckton, 1997. Multiagent Manipulator Control. Ph.D. Thesis, The University of British Columbia,
Department of Mechanical Engineering, Vancouver, B.C.
L. Monostori, B. Kádár, 1999. Agent-based Control of Manufacturing Systems. IPMM'99 -2nd International
Conference on Intelligent Processing and Manufacturing of Materials, Honolulu, HI, 1, 125-131.
 29

L. Monostori, B. Kádár, J. Hornyák, 1998. Approaches to manage changes and uncertainties in manufacturing,
CIRP Annals, 47(1), 365-368.
J. Nagel, J.A. Meech, 1995. Knowledge Accumulation Techniques in MINEX - An Electronic Field Guide to
Rocks and Minerals, IFSA-95, 6th Int. Fuzzy Systems Assoc. World Cong., Sao Paulo, 481-484.
National NEMO Network, University of Connecticut, http://nemonet.uconn.edu/index.htm
G.M. Nicoletti, 1999. Redefining the Web: Toward the Creation of Large-Scale Distributed Applications. 2nd
Inter. Conf. on Intelligent Processing and Manufacturing of Materials, Honolulu, HI, 2, 1013-1018.
E. Niemelä, T. Zhao T. Virvalo, 1994. Fuzzy Logic control in hydraulic mechatronics applications. Intern.
Conf. on Machine Automation (ICMA '94), Mechatronics Spells Profitability, Tampere, Feb, pp. 14.
R. Orfali, D. Harkey, J. Edwards, 1997. Instant CORBA, Wiley Computer Publishing, NY.
A. Papavasileiou, E. Gegov, K. Gavros, N. Sariannidis, 2002. Methods for Decision-Making - Risk
Assessment in the Design of Technological Equipment, Conf. on Mechanization, Electrification and
Automation in Mines, Sofia, Annual of the University of Mining and Geology, 44-45(III), 71-73.
H. Peltonen, 1999. Software toolkit for design of Fuzzy applications in industrial automation,
Prosessitutkimus H. Peltonen Oy.
A. Poanta, D. Dojcsar, 2001. The Fuzzy command of the ventilation process, ICAMC-2001: 14th Intern.
Conf. on Automation in Mining, Tampere.
P.J. Poirier, J.A. Meech, 1993. Using Fuzzy Logic for on-line trend analysis, 2nd IEEE Conference on Control
Applications, 1, 83-86.
V. Polotski, 1995. Guidance of an Articulated Vehicle, Proc. IFAC Workshop on Motion Control, Munich.
J. Puhakka, 1996. Active sludge plant supervision and quality control, Neles Automation Group, Helsinki.
J. Raatikainen, 1998. Raw material management from quarry to kiln [IMA QCS], IMA Engineering Ltd Oy,
Raw Material Management, International Cement Review, 5. www.ima.fi
H. Ranjbar, H. Shahriaria, M. Honarmandb, 2002. Integration of Aster and Airborne Geophysical Data for
Exploration of Copper Mineralization. Case Study of Sar Cheshmeh Area, University of Kerman, Iran.
F. Rosenblatt, 1958. The Perceptron: A Probabilistic Model for Information Storage and Organization in the
Brain, Cornell Aeronautical Laboratory, Psychological Review, 65(6), 386-408.
D.E. Rumelhart and J.L. McClelland, 1986. Parallel Distributed Processing: Explorations in the
Microstructure of Cognition. Volume 1, Foundations, MIT Press, Cambridge, MA, USA.
R. Santos, J.A. Meech and L. Ramos, 1995. Thickener Operations at Carajas Using a Fuzzy Logic Controller,
IFSA-95, 6th Intern. Fuzzy Systems Assoc. World Congress, Sao Paulo, Brazil, 493-496.
A. Satyadas, K. Krishna-Kumar, 1996. Genetic Algorithms and Soft Computing: EFM-based Controllers for
Space Station Attitude Control – Application and Analysis, Physica-Verlag, 152-171.
K. Savolainen, 1998. Higher-level control system for lime kilns, Ahlstrom Machinery Corp., Savonlinna, Finland.
H.D. Scott, 1998. Vulnerability and Use of Ground and Surface Waters in the Southern Mississippi Valley
Region, Arkansas Water Res. Cent., MSC-102, 1997 Ann. Program Rep. Mar 97 - Feb 98.
M.A. Shalan, M.K. Arora, J. Elgy, 2003. CAFCAM: Crisp and Fuzzy Classification Accuracy Measurement
Software, 7th Intern. Conf. on GeoComp., Univ. Southampton, U.K.
B. Shrestha, L. Duckstein, E. Stakhiv, 1996. Fuzzy rule-based modeling of reservoir operation, ASCE J.
Water Res. Planning and Management, 122 (4): 262-269.
T.N. Singh, R. Kanchan, A.K. Verma, K. Saigal, 2005. A comparative study of ANN and Neuro-Fuzzy for
the prediction of dynamic constant of rockmass, J. Earth Syst. Sci., 114(1), 75{86}.
P.K. Singh, 1998. Application of Fuzzy Logic in Rock Mechanics, ROCKSITE, Nagpur University, India.
F.L. Smidth Group, 2003. ECS/Fuzzy Expert SAG Mill Control, FLS Automation Co. brochure, pp.4.
http://www.flsmidth.com/flsmidth+automation/english/contact/ brochures
+and+info/product+profiles/ecs/fuzzyexsagmillcontrol0.pdf#Page=3
F.W. Su, 1998. Dephosphorization of magnetite fines: surface reactions, flotation kinetics and Fuzzy Logic
application, Doctoral Thesis, Chemical and Metallurgical Eng., Lulea University.
J. Sulkanen, R. Tuokko, 1994. Towards Effective Mining – Tele-operation in the Mine, Regional APCOM
Symp. on Comp. Appl. in the Mineral Industries, Bled, Slovenia, June, pp. 8.
J. Sutherland, 2002. Holographic Neural Technology: Systems and Applications. in: Intelligence in a Materials
World: Eds. J. Meech, M. Veiga, Y. Kawazoe, S. LeClair, CRC Press, Boca Raton, FL, 93-115.
E. Szczerbicki, Z. Gomolka, 1999. Management of Information in Complex Systems: Perspectives for the
New Millennium. IPMM'99 -2nd International Conference on Intelligent Processing and Manufacturing
of Materials, Honolulu, HI., 2, 749-752.
 30

V.M. Torres, A. Chaves, J.A. Meech, 2000. Intelligold - an Expert System for Gold Plant Process Design,
Cybernetics and Systems, 31(5), 2000, 591-610.
R. Tuokko, J. Sulkanen, T. Helgesson, 1994. Long distance telerobotics using wide-area-networks, Intern.
Conf. on Machine Automation (ICMA '94), Mechatronics Spells Profitability, Tampere, pp. 11.
P. Valckenaers, F. Bonneville, H. van Brussel, L. Bongaerts, J. Wyns, 1994. Results of the holonic system
benchmark at KULeuven, Proc. of 4th International Conference on Computer Integrated Manufacturing
and Automation Technology, Troy, New York, 128-133.
H. van Brussel, J. Wyns, P. Valckenaers, L. Bongaerts, P. Peeters, 1998. Reference architecture for holonic
manufacturing systems, Computers in Industry, Special Issue on IMS, 37(3), 255-276.
M. Veiga, J. Meech, 1995a. A Fuzzy Model for Risk Assessment. Proc. IEEE SMC, Vancouver, 1640-1643.
M. Veiga, J. Meech, 1995b. An Adaptive Fuzzy Model for Risk Assessment of Mercury Pollution in the
Amazon. Proc. 6th Int. Fuzzy Systems Assoc. World Congress, São Paulo, 1, 489-492.
M. Veiga, J. Meech, 1995c. HgEx - A Heuristic System on Mercury Pollution in the Amazon. Water, Air &
Soil Pollution , 80, 123-132.
M. Veiga, J. Meech and D. Tromans, 1995. Mercury Emissions and Stability in the Amazon Region, in
Waste Processing and Recycling in Minerals and Metallurgical Industries II, Eds. S. Rao, L.
Amaratunga, G. Richards, P. Kondos, (Eds.), 34th Conf. of Metallurgist of CIM, Vancouver, 523-538.
M. Veiga, J. Meech, 1997. Fuzzy-neural Systems for Adaptive Reasoning on Environmental Risk Analysis.
MineIT '97, 1st Intern. Symp. on Information Technologies in the Mineral Industry via the Internet, G.
Panagiotou, Ed., Dec., CyberSpace and Athens, Greece, 10 HTML pages.
F.Y. Wang, P. Lever, 1994. Automated Robotic Mining Excavation Using Fuzzy Logic and Neural
Networks, University of Arizona, Tucson, AZ, NSF Award Number – 9411810.
T. Waters, A. Evans, 2003. Tools for web-based GIS mapping of "Fuzzy" vernacular geography, Proc. 7th
Intern. Conf. on GeoComputation, Univ. Southampton, U.K.
L. While, L. Barone, P. Hingston, S. Huband, D. Tuppurainen, R. Bearman, 2004. A Multi-Objective
Evolutionary Algorithm Approach for Crusher Optimisation and Flowsheet Design, Comminution '04.
K.W. Wong, Y.S. Ong, H. Eren, C.C. Fung, 2004. Hybrid Fuzzy modelling using a mimetic algorithm for
hydrocyclone control, Proc. 3rd Intern. Conf. on Machine Learning and Cybernetics, Shanghai, Aug.
J. Yen, 1999. Fuzzy Logic-A Modern Perspective, Knowledge and Data Engineering, 11(1) 153-165
L. Yliniemi, 2001. Adaptive Fuzzy Control of a Rotary Dryer, In: K. Leiviskä, (Ed.), Industrial Applications of
Soft Computing in the Paper, Mineral & Metal Processing Industries. Physica-Verlag, Heidelberg, 109-132.
L. Yliniemi, J. Koskinen, K. Leiviskä, 2003. Data-driven Fuzzy modelling of a rotary dryer. Intern. J. of
Systems Science, 34(14-15). Taylor & Francis, UK, 819-836.
L.A. Zadeh, 1965. Fuzzy Sets, Information and Control, 8, 338-355.
T.Q. Zeng, P. Cowell and Q. Zhou, 1997. The Usher's Approach (UshA) for Realisation of Fuzzy Clustering
Theory for Zonal Analysis in Raster-Based GIS, Proc. 2nd Intern. Conf. on GeoComputation, University
of Otago, NZ, Aug. 26-29, CD-ROM.
A.X. Zhu, B. Hudson, J. E. Burt, K. Lubich, 2001. Soil mapping using GIS, Expert Knowledge and Fuzzy
Logic, Soil Science Society of America, 65, 1463-1472.
A.X. Zhu, R.X. Wang, J.P. Qiao, Y.B. Chen, Q.G. Cai, C.H. Zhou, 2004. Mapping landslide susceptibility in
the Three Gorge Area, China using GIS, Expert Systems and Fuzzy Logic: in: GIS and Remote Sensing
in Hydrology, Water Resources and Environment, IAHS Publication 289, Intern. Assoc. of Hydrological
Sciences, Wallingford, UK, 385-391.
A.X. Zhu, 2005. Research issues on uncertainty in geographic data and GIS-based analysis, in: R. McMaster,
L. Ursery (Eds.) A Research Agenda for Geographical Information Science, CRC Press, NY, 197-223.