Top of page
ABSTRACT
Over the past 15 years, Artificial Intelligence techniques have shown promise of providing the ability to capture "human intelligence" into a computer program. The sub-fields of Expert Systems and Artificial Neural Networks are finding increased use in industry to perform tasks previously considered the domain of human thinking. But, are these methods just a new type of software much like that of the past or is real magic hidden within the lines of code? Will this field continue to enjoy high growth rates and if so, what developments will we see in the near future? Which algorithms and methodologies will continue to grow?
This paper discusses a number of paradigms which can be used to create successful applications. There are certain barriers; some psychological -- some linked with knowledge acquisition, that must be overcome or avoided to ensure system performance. Development time also requires careful consideration before embarking on an AI project. The paper focuses on constructing goal-driven systems that transfer high-technology solutions from the research laboratory to the plant floor.
Respect for system users is of paramount importance -- whether their involvement is to access contained information or to participate in program maintenance as new knowledge becomes available. Examples of successful and unsuccessful AI implementation in Canadian industry and academia provide evidence of the importance of this issue. The paper also examines some current constraints on applying AI in the real-world. Among these problems are automated knowledge acquisition, methods to acquire data during run-time, real-time systems (hardware versus software), explanation and justification features and adaptive/learning systems.
INTRODUCTION
Artificial Intelligence has experienced rapid growth over the past decade. Numerous Expert System development tools entered the market-place in parallel with the explosion of microcomputers and there are many examples of successful applications. Processing plants worldwide, now use AI to solve real-world problems.1-10 Early methodologies have distilled into rule-based environments with techniques that link incoming data to sub-goals and eventually, to final conclusions. Frame-based approaches have evolved into object-oriented methods that group information into classes allowing construction of very efficient rule-structures in terms of both reduced numbers of rules and processing time.
But, as more and more technical people have become familiar with these methods, it is apparent there are limitations to the ease with which applications are developed. It takes considerable skill in the following areas to ensure success: psychology, computer programming, knowledge acquisition and representation, etc. Without training, it is difficult to build useful and productive systems - ones with big payback!
Expert Systems have promised to embody "human intelligence" into a computer program. This they do very well and a system can function with only a single rule contained in its knowledge base making development relatively painless. Such systems have limited ability and knowledge but not necessarily all. There are techniques today to incorporate expertise directly into a Hypertext Interface so Users control and manipulate the domain as they desire.7
Early systems were considered small or toy-like unless they contained hundreds or even, thousands of rules. Today, a system containing this number of rules is either an extremely large application with extensive domain-specific knowledge and perhaps, some global-knowledge capability, or it has been built using old techniques. Computationally-intelligent methods such as Fuzzy Logic, Artificial Neural Networks and Evolutionary Computing significantly enhance the ability to represent knowledge effectively and efficiently.
This paper describes the features of AI techniques that can be used to create productive systems. We will begin by examining components of human intelligence that are currently well-modeled. The approach to building a successful system is then presented. The drawbacks of existing methods and the attempts to address these issues are discussed together with the introduction of a new AI field - Computational Intelligence. Several successful systems are used throughout the text to illustrate the methods indicated.
ELEMENTS OF HUMAN INTELLIGENCE
To understand how we can simulate the human-thought process, it is necessary to examine some components of our thinking ability. A partial list might include the following items:
Current Expert Systems handle some of these but are poor at mathematical problem-solving (too-slow); they cannot learn and are usually not adaptable; generalization and summarization skills are essentially non-existent. In real-time, when properly configured, many measurement-reaction situations that require at least several seconds turnaround time can be handled, but faster times are still unavailable. Speed issues will eventually be addressed with more powerful hardware such as parallel-processing computers but this is several years away.
Multiple output is essential for these systems but not necessarily in the way humans communicate. A User Interface should provide graphical presentation of data that is easy to comprehend. Access to additional background data should be available. Most AI tools today can communicate with other software, either directly through RAM, disk-stored data files or across a series of networked computers (the first use of parallel-processing devices).
Data entry by keyboard and mouse are the best way to communicate with an Expert System. Audio alarms are useful to draw attention to key process upsets but voice-activated messages and voice-recognition capability are still not advanced enough. In a few years however, this situation is likely to change. Already, there is software on the market that provides dictation ability and voice command-entry. Some plants have tried on-line audio/visual help menus but these quickly become repetitious and are only beneficial the first few times they are used. Development time is too long to justify the small rewards. Such multi-media features are too static and require too much storage space to be useful in the long run -yet!
The key area of success with this technology is the ability to rapidly model a process with almost the exact terminology used by an Expert. The features of knowledge representation in a Development Tool can be learned quickly but the time to acquire knowledge is not short as the system must be built one step at a time - acquiring each jewel of wisdom bit by bit.
REPRESENTING HUMAN INTELLIGENCE
Expert Systems possess the ability to embody "intelligence" into a computer program in much the way human communication takes place. Facts are stored symbolically with their truth or falsity dependent upon current circumstances while using a system. A Knowledge Engineer can build a system using the exact words of the Expert expressed as symbols and rules-of-thumb that link these symbols. For example, an Operator might express the following idea:
"When froth conditions on the lead rougher bank are porridge-popping, I reduce the collector addition rate by at least 100 cc/min unless it is already at minimum. In that case, I increase the water addition to the rougher feed provided the Pb feed grade is not high. If it is high, then I reduce the feed tonnage."
This simple paragraph can be formulated into the following code:
Rule Porridge-popping
IF froth.conditions.porridge-popping
AND collector.addition_rate.minimum is FALSE
THEN collector.addition_rate_change.Negative_Big is TRUE
ELSE MACRO ("Water_addition_increase")
Rule Water_addition_increase
IF froth.conditions.porridge-popping
AND collector.addition_rate.minimum
AND feed.Pb_grade.high is FALSE
THEN feed.water_addition.Increase is TRUE
ELSE feed.tonnage_rate.Decrease is TRUE
Backing-up these rules would be Fuzzy Sets to map discrete measurements into the linguistic names used to create the rule statements. Explanations can be attached to each fact and rule descriptions to each rule as desired to provide Users with the ability to probe the knowledge base for further details should these rules fire successfully.
These rules and others are added to the system incrementally with static and dynamic testing of the knowledge base at various points in time as the system grows. Other "experts" might be polled to ensure there is general agreement about the strategy to be implemented in the system. If conflict arises, a method to arbitrate disagreement must be setup, usually with the support of suitable metallurgical staff.
During a consultation with an Expert System, a User can obtain justification for the advice given or ask the system about the meaning of certain dialog: "Why are you asking me for this information?" or "What does that term mean?" or "How did you arrive at your conclusion?" In this way, such systems serve an important training function in addition to fulfilling their original goal: diagnosis, monitoring or control.
While this provides a dialog which mimics how a real "Expert" might communicate expertise to a novice or lesser expert, there is more to human intelligence than providing advice and explanations. For example, these systems cannot learn from experience. Neither can they devise new knowledge from an examination of their current knowledge. They simply perform what a human has created them to do. So based on today's techniques, Expert Systems appear to have reached a watershed.
PROBLEMS WITH CURRENT AI TECHNIQUES
Domain Selection
The choice of an appropriate domain to begin applying AI is not always clear. Often, a careful organization selects an application that is easy to implement in order to demonstrate to skeptical individuals that the technology does work. A consultative system is generally the preferred starting point with development of an off-line advising system. Once the technology is proven, other applications should grow naturally out of this initial success. While this conservative approach is likely to overcome resistance to AI, it is a slow route and does not produce large rewards quickly.
An alternate approach considered more risky and certainly more costly, is to begin with an on-line real-time monitoring or control system. These are high-payback applications which will pay for their initial cost in several weeks or months. The initial prototype is developed fast. Select a well-understood problem area; one that is difficult to automate but has existing sensor data. Build a system with a useful and user-friendly interface so operating personnel adopt the system as an essential took in their work.
On-line systems can be prototyped in as little as 2 weeks by experienced personnel. With close cooperation of Management and Users, these systems produce major benefits9 in as little as one month. The process begins by rapid development of the knowledge base and User Interface. Once tested in a static mode, the system is hooked into the plant DCS database or PLC system to operate in a monitoring mode in parallel with existing manual control. System advice or diagnostics are evaluated by the Users together with the development team and modified to suit the objectives. Once all parties agree, the system is turned on to control mode to supervise existing controllers with set point changes.
Development Time
The time required to develop an Expert System is a function of a number of variables:
The first two items can be met with a good training program and acceptance of the technology by a variety of levels in an organization. Commitment to the project and the technology are essential to initiate successful development. However, while learning the tool and techniques can be accomplished by interested people in a very short time (2-3 days), the time to apply these methods to a knowledge domain can be considerable. Mastery of these techniques is a life-time commitment as most tools today possess a multitude of approaches to constructing and organizing a knowledge base - meaning there is always a better way to build a system.
The biggest roadblock to success is acquisition of knowledge from an Expert. Our experience has been that using a person other than the Expert for development is a better approach. The "Knowledge Engineer" should not be the domain Expert but must be motivated to become one and have more than just a passing interest in the subject. Development of the system occurs using an interview process between the Knowledge Engineer and the Expert. Experts often have great difficulty describing stepwise how they apply their knowledge to make decisions. Analysis of situations and cases can often be frustrating. A Knowledge Engineer must be patient and prepared to ask "silly" questions that often open up new paths through the domain in moving from input data through to final conclusions.
A sensible framework2,11 for questions to pose during the interview process might be:
Meta-knowledge such as explanations and rule descriptions can also be important in establishing future avenues to be opened up later during further interview sessions. Interviews should be scheduled and of short duration (1-1.5 hours). Interruptions should be avoided and all parties should respect each other and have commitment to reach the planned goals.
System Maintenance
Once a system has been installed, it is important to continue development and expansion of the knowledge base. New knowledge may become available during the construction process or during the early use of the installed prototype. As the system is used, knowledge may require modification with deletion of certain concepts and addition of new features.
To accomplish this task, an individual "champion" of the project within the organization is required to monitor system progress and adjust the knowledge base when unforeseen occurrences are encountered. Training this person in the "tricks" and "techniques" of the tool is essential to continue the process of transferring AI methods into the organization.
Learning Techniques and Adaptation
Current Expert Systems do not possess true learning capability. They are unable to adjust their expertise to account for error or new happenings. They cannot be introspective and examine what they currently know in order to propose new relationships or ideas. However, there are a number of Development Tools that possess techniques for dealing with uncertainty that allow a system to adapt its rule structure as circumstances change without adding new lines of code. Once an input/output map has been constructed linking certain inputs to one or more outputs, it is possible to incorporate a new variable into the model without significantly adjusting the code. The methodology that permits this adaptation is called Fuzzy Logic.12
An example of fuzzy set adaptation is as follows:
Suppose we have mapped the variable flocculant.addition_rate.@f onto a linguistic expression such as flocculant.addition.low as below:
Fuzzy Set: addition_low
Source: flocculant.addition_rate.@f
Now suppose the slimes content of the feed slurry becomes very-high. Under this circumstance, a flocculant addition rate of 30 ml/min is no longer "not low"; instead we find that as much as 60 ml/min might be considered low.
To accommodate this adaptation, our fuzzy set can be easily changed to:
Fuzzy Set: addition_low
Source: flocculant.addition_rate.@f -
0.03 * CERTAINTY(feed.slimes_content.very_high)
Now belief in a low flocculant addition as a function of the actual addition rate becomes a fuzzy concept with respect to the feed slimes content. When the slimes content is definitely very high the fuzzy set shifts to full belief at 30 and full disbelief at 60. But when the slimes content is definitely not high, the original definition applies.
There are many ways in which adaptation of a "fuzzy" linguistic concept can be implemented. Any mathematical relationship can be handled through modification of the Fuzzy Set source. Adaptation is an important way to expand an existing system without disrupting the rule structure or changing the original logic in the system. Although not strictly a "learning" technique, the relationship originally established between input and output variables is adjusted in response to changes in variables external to the original rule structure.
IMPLEMENTATION PROBLEMS
Implementation of an Expert System is perhaps the most difficult stage in the development cycle since it involves so many different people and ideas. Success demands excellent interpersonal skills with a little dab of psychology and problem anticipation thrown-in for good measure. We are talking here about introducing a system into a large organizational network or into a plant to conduct process monitoring and control. Some of the important issues to be addressed during implementation are as follows:7
EXTERNAL ENVIRONMENT
DATA FACTORS
Obtaining system data involves decisions about sources, sensors and time. Data problems can railroad a system very quickly. For real-time applications, the speed at which information is processed must be considered. Data pre-processing is required to ensure rapid reaction. This requires application of intelligent drivers to interface with the System and the Process Sensors to filter data in a "smart" way producing symbolic meaning for the system's knowledge base.
PROCESS AND PROJECT FACTORS
TECHNICAL FACTORS
PEOPLE FACTORS
Interpersonal rivalries can prevent success. If the Developer is responsible for implementation, strong persuasion skills are needed to work with Management and Users. If Management or Users are to conduct implementation, then excellent communication is necessary to ensure the project needs are addressed by the Developer. If only the User Group is responsible for implementation, then the initial stages of the project are extremely important to ensure proper training is given about installation and maintenance. The Development and Implementation cycles must be continuous and carefully planned if all parties are not involved in installation. An external consultant is often useful to begin the introduction of AI into an organization.
When conditions that promote successful implementation are not addressed, some Users and/or Management personnel may develop tactics for "counter-implementation" or sabotage.
These include the following:
Changing goals is the simplest way to sideline a project, so be sure that all parties agree on clear and concise objectives at system initiation. Whenever this tactic is tried, redirect attention to the initial goals document and show how the changed goal will impede timely completion within budget. There are many ways to address such tactics but it will be a difficult up-hill battle without support from Management. Remember - resistance to change is a normal "human" feeling about new ideas that derives naturally from one or more of the following:
The key is User commitment. Here are some ways to get this involvement:
Lack of commitment is a major cause of system rejection. Users must believe ES technology is useful and can increase knowledge. Often, thoughtless postures are taken: "I didn't try it because I didn't like it". One example of failure caused by this issue is the Polaris Expert7,8 - the first Expert System for real-time supervisory control in the Mining Industry.
Located 150 km from the North Pole, the Polaris mine appeared a perfect place to introduce this technology to standardize operating practices, stabilize circuit conditions, provide training for inexperienced operators (turn-over was high at Polaris), and guide plant operators. The system developed in concert with the operators over a one year period. Good performance is claimed but the system needed regular maintenance. Failure occurred because the "champion" left the company and commitment to maintain the system was low. The system worked well as long as a "Baby-Sitter" was available. It seems strange that a company would spend large sums of money on a project and then let it die because no one was available to keep it going.7
Political problems are also sources of failures. Sometimes, union personnel do not understand that consultative ES are useful for training novices and are not aimed at worker replacement. This creates a very uncomfortable situation and an ES user can become a traitor to his peers. When attempting to capture knowledge from operator experience, any suspicious attitude can destroy the trust that must exist between operators and knowledge engineer.
The flotation advisory system built at Highland Valley Copper, Logan Lake, BC, exemplifies the reluctance of older operators to share their expertise with a knowledge engineer. Some had little faith in the ES technology for training younger operators and refused to contribute to the project. When the knowledge was finally obtained and the system implemented, relief-operators responded positively to the new ideas and actions provided.9,10
To attack resistance to change, three stages can be used by system developer and users:
MELT: To create an awareness of the need for change at the organizational level or for individual workers, incentives are a powerful persuasion technique. These can include: increased bonus for high productivity, decreased "hard" work / increased "smart" work, better working conditions, more free-time to perform routine or lower-priority jobs, etc.
SHIFT: Next, forces that impede introducing change are identified, diminished in intensity or redirected to develop new methods, learning attitudes and behavior. The style of implementation will be important. User involvement in construction is best at this stage. Don't force the system onto a hostile group. Introduce training sessions early.
CRYSTALLIZE: The final step reinforces the changes to maintain and stabilize the situation at a new level. Allow Users to evaluate and test the system - their ideas must be listened to and considered. Set up Daily Report sheets to be filled-in. These documents should be easy to draft with minimum effort and time. Users should believe the system is an extension of themselves and they are the custodians of the knowledge being utilized. If the system becomes essential to the new routine, complete acceptance is assured.
REAL-TIME PROCESSING
Rakocevic et al13 consider the following definition for a real-time system:
An integrated computer program environment that responds fast enough to interrupting external events to provide accurate and fast control (or alarm) action.
At the lowest level in the process control hierarchy - Level 0, instruments monitor, sense and manipulate plant variables. These devices are connected to control units able to implement control logic - Level 1, consisting of Programmable Logic Controllers (PLCs), single-loop controllers or Distributed Control Systems (DCS). Collection, manipulation and presentation of sensor data are based on numerical methods. At this level, a control system responds extremely fast to external events, carrying out several activities simultaneously. To provide such speed on a computer, a real-time multi-tasking Operating System is a necessity. Most AI applications for real-time control are implemented at the Supervisory stage - Level 2, although there are a number of examples of non-real-time applications at higher levels for diagnosis and advice - Plant-Wide and Total Enterprise Levels.1,6,7,
Figure 1. ProcessVision - A Real-Time Multi-Tasking Intelligent SCADA Package.
The ProcessVision software package depicted in Figure 1 is an Expert System-based SCADA tool for Supervisory Control.11 Cycle-time through the Knowledge Base is typically several seconds. Rakocevic et al7 refers to this time scale as "pseudo" real-time. Similar to a real-time system, "pseudo" real time provides fast and accurate control, but in a different time domain.
AI will have increased success with development of parallel-processors and faster computers. More powerful CPUs within multi-processing environments can solve many of the speed issues allowing conventional AI to be used in real-time. Meanwhile, instead of waiting for the big event, people are studying ways to apply AI in real-time using software approaches. Hardware does exists to solve some AI difficulties, but data acquisition boards with on-board CPU chips and/or programmable EPROMs are not essential. We can provide these functions today using software on high-speed single CPU machines (486-66, -100 or Pentiums)
COMPUTATIONAL INTELLIGENCE - A NEW APPROACH TO AI
Processing symbols on conventional "von Neumann"-type computers is inherently inefficient because of hardware architecture. Successful AI applications focus directly on manipulating symbols. So how can AI move effectively into an environment (real-time, direct control, etc.) that demands intensive and rapid numerical computing? Part of the problem relates to combinatorial complexity. As the number of inputs and outputs in a system increases, difficulty in dealing with all combinations of these occurrences increases exponentially.
Symbolic methods, while useful at providing explanations and access to the inner workings of a system, are clumsy and onerous to develop and maintain but algorithms are available to support the high overhead of symbolic processing on a microcomputer. Without knowledge of these methods, ES applications will remain trivial, slow to develop and difficult to maintain.
The answer lies in a set of paradigms called "Computational Intelligence" that have evolved out of AI and other related disciplines. Computational Intelligence (CI) is a term coined by Bezdek in 1993 to describe "low level" knowledge in the style of the mind.14 CI consists of very "primitive" concepts in the AI sense, that support the beginnings of Symbolic Knowledge which Bezdek called "tid-bits". These "elements" can become the inputs to an AI structure that processes symbols heuristically or in other ways. Primitives are fundamental operations conducted on numbers:
These fundamental operations are the primitive steps in any complex numerical structure that produce elementary symbols for AI processing. Current hardware can deal with these numerical manipulations efficiently suggesting that CI can become the fundamental support structure for conventional AI methodologies. The components of CI may include Fuzzy Logic mappings, Artificial Neural Network connections and/or Genetic Algorithm optimizations - all of which depend on numerical processing to support the generation of symbolic output.
This definition seems limiting in that Computational Intelligence should not rely only on pure mathematics. For true intelligence, CI should include AI methods such as heuristics to allow rapid computations, even when gross error may be contained in the output.
To bring AI into the control hierarchy of a real-time environment, a CI module that creates "primitive" symbols has to be very fast (~ millisecond processing time).13 The key trade-off in all real-time computing systems is always accuracy versus processing speed. Incorporating AI methods into a CI module introduces certain error, but produces appreciable speed. With proper setup, direct connection between the CI and AI levels can assist with error detection and interpretation at the AI level. The AI module can test symbolic outputs from co-operating sensors, and tune out the error using a feedback loop to the CI module as in Figure 2.
Figure 2. Error Detection in the AI module to adapt the CI module.
Borrowing an analogy from Bezdek,14 an individual given an apple can immediately recognize its shape and colour. With eyes closed however, other senses must be used: touch, smell and taste. Smell and taste produce real-time decisions about the apple but touch and feel may require additional processing to visualize mentally an apple's shape and texture. Such symbols must be recalled and compared with measurements less accurate than sight. So too can CI assist AI to make rapid decisions intelligently, monitor the selection and correct it if necessary.
To do this effectively, CI need the following items:
The approach is similar to the hierarchy of human intelligence depicted in Figure 3. Biological Intelligence consists of manipulating symbols supported by low-level numerical processing to generate belief in a particular symbol. This hierarchy is mirrored in the arrangement of AI with Computational Intelligence forming the foundation for rapid problem-analysis.
Figure 3. The Similarities of Biological and Artificial Intelligence
The boundary between AI and CI is "fuzzy", with Fuzzy Logic, Artificial Neural Networks and Genetic Algorithms playing a cross-over role. In the future, the boundary between Biological and Artificial Intelligence will also be "fuzzy" as robots or clones indistinguishable from humans become common. Machines will have chemically-based processors; people will have electronic-thinking replacement parts but this is well into the future.
If we apply crisp concepts to define the components of intelligence, we end up with 3 distinct levels: Biological Intelligence (BI) or Artificial Intelligence (AI), Symbolic Intelligence and Computational Intelligence (CI). In a similar fashion to man-made intelligent machines, we also find different types of Biological Intelligence (otherwise we might not eat beef!). Humans are at the highest level - the "sentient" species. More primitive forms of life are less intelligent and yet perform complex tasks effectively and efficiently.
Much AI work has attempted to imitate the minds of certain organisms. For example, DARPA set a goal in 1987 to mimic the intelligence of the honey bee by the year 1992.7 This requires 109 connections in an artificial neural net. We do not know if the project succeeded, but this work is moving us toward Artificial Life and Virtual Reality. So, is there a boundary between AI and BI, between a true reality and a virtual one? CI is an element of this boundary.
Within Biological Intelligence, the ability of autistic savants to carry out rapid and accurate data calculations, musical recall, etc., are examples of how the human brain can perform accurate real-time computation. Whether output is intelligent or not is determined by those who interact with such exceptional people. Perhaps they use fuzzy set definitions with very broad support characteristics (see Figure 4). Such curves provide rapid interpolation, but sensible output from an individual set is not clear. While savants can tell the day of the week for a particular date, they are unable to understand the significance of the date in question.
Figure 4. Examples of Fuzzy Set definitions for normal and autistic humans.
FUZZY SYSTEMS
Fuzzy Systems are Expert Systems that use Fuzzy Sets and Fuzzy Logic to perform qualitative modelling and/or to control processes when no model has yet been identified or developed from fundamental principles. Fuzzy Logic allows rules to generalize across a space-map by interpolating from rule node to rule node. While the answer is an approximation of the "correct" one, it is usually acceptable. Gains in search and storage efficiencies are remarkable.
Fuzzy Systems are adaptable, in that mapping of fuzzy concepts on a Universe of Discourse can change in response to external forces. The context of the Fuzzy I/O Mapping (often called a Fuzzy Associative Memory or FAM) can be taken into account so that the term "tall" for example, can mean one thing to a pygmy and another to a giant.
A Fuzzy Logic Controller15 operates with a series of rules that link input linguistic expressions to output expressions. The rules combine inputs using AND and OR operators. An Inference Method is used to determine the Degree of Belief in each output expression. Typically, the Minimum degree of belief is selected from ANDed facts and the Maximum degree of belief from ORed facts. This generates a Net Degree of Truth in each output expression which is then multiplied as a fraction by the Fuzzy Set definition for each output. To calculate a discrete output value, the individual output Fuzzy Sets are combined into a single distribution function using a weighted-average approach. This function is then defuzzified by taking the centroid position. Smith16 has delineated over 90 different Inference-Defuzzification options and has shown the advantage of switching methods dynamically when circumstances change.17
Alternatively, each single output Fuzzy Set can be "linguistically"
defuzzified18 into a variety of expressions based on the degree
of belief in the concept. Adaptation of these expressions is also
possible by using an "alpha" or acceptance factor. This
permits generation of a multi-variable FAM that generates variable
"fuzzy" output expressions. Figure 5 shows an example
of how the "acceptance" factor can be determined from
three main types of variables - in this case Economic, Technical
and Social-Political variables are combined to generate "alpha".19,20
Figure 5. A 3-dimensional Fuzzy Associative Map used
to calculate an Acceptance
Factor (alphaHEF) to adapt output from a Fuzzy Logic Controller.
Each variable grouping derives from many external variables that use other FAM rules. This technique was devised to accommodatedifferences in the perception of the intensity of man-made mercury emissions in a large Expert System designed to diagnose Hg-pollution problems in the Amazon.18 It was considered that behavior of workers depends on society incentives and reactions.20,21 As a result, developments in a society change our view of high and low levels of Hg emissions. A high alphaHEF indicates acceptance of Hg use and low control of emissions by a society, country or region. For the Amazon, alphaHEF is 1.0, while for Canada, where Hg is practically banned and well-monitored by authorities, it is much lower (0.1 or 0.01). For Canada 150 years ago, when the hazardous effects of Hg were unknown and thousand of miners were colonizing the West, alphaHEF would be much higher (10 or 100).
A term called "High Emission Factor" is used to represent the collection of observations about a mining operation that derive belief in high emissions. Belief in HEF is linguistically defuzzified according to the relationship depicted in Figure 6. Note how the term used to describe Hg emissions is dependent on the value of the alphaHEF. When alphaHEF is high indicating acceptance of Hg use, to conclude that emissions are "extremely-high", belief in HEF must also be very-high (>98%). When alphaHEF is very-low indicating rejection of Hg use, "high" emissions derive from belief in HEF as low as 7 %. Elasticity with this technique is substantial.
ARTIFICIAL NEURAL NETWORKS
The area of Computational Intelligence offering true learning capability is that of Artificial Neural Networks. The paradigms in this field are based on direct modelling of the human neuronal system and hence, researchers in this area are often referred to as "Connectionists".22
Basically, an I/O space map is modeled as a series of interacting nodes in a network. When a node fires (taking on a signal value > 0), the signal is amplified by multiplying by a "weight" assigned to each connection between this node and others. At the receiving node, all incoming signals are summed to give an overall force acting on the neuron. This sum is passed through an activation function to set a signal strength from 0 to 1 for the neuron. This signal is sent to other nodes to which this node is connected. The process continues with a new node until all nodes are exhausted and the signal strength of all outputs are known.
The full power of this method will only be exploited when parallel-processing computers become more widely available so that many neurons in the network are dealt with at once. Nevertheless, implementation of these systems with large numbers of connections (hundreds) perform extremely well on today's high-speed microcomputers.22,23,24 Figure 7 shows the structure of a 6x6x6 3-layer network with bias.25
Figure 7. Structure of a 3-Layer Artificial Neural Network with Bias.
The advantage of neural networks is their ability to learn or adapt to changing circumstances. Learning begins by selecting a set of random weights and presenting the network with a set of known I/O values from a large set of training data. The network is fired and predicted outputs are calculated. These are compared with the desired output and the error passed back through the network by using a regression or gradient-descent technique. New weights are derived and the network examines a new set of I/O data. This process continues until the overall error is within a predefined tolerance or until a selected number of iterations have executed. This is known as Supervised Learning - Back-Propagation being the most widely-used technique.23
The trained network is put into service to deal with the particular environment it was designed to handle. Its performance is monitored on a regular basis by examining the accuracy of its predictions. Should drift occur due to changing external forces, the network can be removed temporarily from service to be retrained on more up-to-date data.
There are a number of ways in which a network can be configured - the Perceptron model, the 3-layer Back-Propagation model, Kohonen networks with unsupervised learning, CMAC or Cerebellum Model Articulation Control, Holographic networks, etc. Each has its own unique advantages and disadvantages. The Perceptron model for instance, cannot deal with problems that have non-separable space mappings. Kohonen networks are used to classify input patterns rather than for process control, etc. Back-propagation methods take considerable time to learn and can get trapped in local-minima regions.7
AND Corporation of Hamilton, Ontario has developed a novel approach to neural modeling very different from conventional techniques. Instead of training a network by feeding one set of I/O data at a time, all data is analysed together to produce an I/O map that directly links input and output data. The method is not a connectionist model. Instead, complex number arithmetic is used to perform regression analysis. The phase angle represents the actual value of the item, while the amplitude of the vector represents its Degree of Belief.26
Holographic systems27 require only single cells to learn stimulus-response associations. These systems can actually learn in real time unlike connectionist models which require considerable time to learn an I/O map. Holographic nets map all input/output data on one pass directly into the structure of the neuron "cell". Speed of learning is orders-of-magnitude greater than existing back-propagation techniques.
Holography is not simply retrieval of patterns. All I/O patterns are actually superimposed onto the same set of synapses. The model has the advantage of conventional network accuracy but at only a fraction of the memory and processing time to perform a similar pattern-ID task. HNet generalizes by performing interpolation among the trained I/O mappings much as occurs in a Fuzzy Logic controller. Similar to the numerous Defuzzification and Inferencing options available, holographic networks allow a User to modify the generalization characteristics so that mappings overlap as interpolation is performed.26
GENETIC ALGORITHMS AND EVOLUTIONARY COMPUTING
There has been much excitement in recent years in the study of Genetic Algorithms - a concept based on Darwin's criterion of "Survival of the Fittest". Genetic Algorithms and/or Genetic Computing are methodologies that automate the search for a better solution. The process involves representing solutions by a string of numbers and "mating" them two at a time. The outcome of this simulated "procreation" has better characteristics than either parent.
The method is an adaptive parallel-search strategy to locate a global optimum point without becoming trapped within local minima positions. The algorithm uses the following operations:
GA systems analyse and evaluate each member of a population of solutions ranking them on a "best-fit" criteria. A number of genetic operators are applied to the population to create a set of new solutions, which are then explored. The process continues until all members converge to a stable position. Three types of operations are used: reproduction, cross-over and mutation. This latter provides a periodic push to find a new and better solution.28
INTEGRATION OF THESE TECHNOLOGIES
To develop learning capability, Expert Systems (whether Fuzzy or not) need to be able to interact with a number of these important algorithms. In the case of the connectionist models, the weights of the network can be converted into rules for incorporation into a Knowledge Base. By examining the weights in a scaled fashion, it is possible to formulate symbolic meaning for the hidden neurons within the structure of the network. In this way the network can be adapted to provide explanations to a User rather than remaining a simple "black-box".
Learning can be accelerated by incorporating a Holographic approach to calculate the weights of the connectionist network. Although direct application of the holographic network is likely the fastest and most efficient process, it too suffers from being a "black-box". The Expert System which polices and controls all learning systems thus, becomes the reservoir for the intelligent symbolic representation of the knowledge domain with its symbols derived from the neural methods which learned from the training data sets.
Development Of Computationally-Intelligent Real-Time Systems
For proper organization, it is necessary to divide the important processing functions into separate modules as shown in Figure 1. Each module coordinates individual tasks such as: data storage, message transfer, alarm sensing, event scheduling, process/user communication, knowledge processing, explaining, etc. The modules interact with each other, the user and the process. As the system grows in size and complexity, availability and turn-around of resources decline until eventually, additional hardware is required (more memory or more CPU).
To delay onset of such problems for an expanding system or to design for maximum efficiency with currently available resources, intelligent computation must be included in the system at the initial development stage. Traditionally, intelligent SCADA systems have been applied for supervisory control service only. Typical cycle times for knowledge base processing average in seconds. Typical data up-date times are 1 second to ensure efficient utilization of system resources. For some applications however, this is too long a delay. Certainly, for direct regulatory control, this sampling interval is inappropriate. Also for rapidly-changing variables that indicate important process transformations, much faster measurements are required.
To apply a rule-base system to batch data such as on-stream XRF assays, temperature profiles of steel billets29 or thickness measurements of paper on a roll from a paper mill, it is necessary to sample at much faster rates. Suppose the temperature of a billet surface is to be recorded as it passes in front of an optical pyrometer. The duration of measurement is relatively short (~ 2 seconds) but the intensity of data output is substantial (perhaps as high as 2000 points). The time interval must drop to 1 millisecond to accomplish this task.
Storage of such data (much of which is redundant) within a conventional SCADA database is inefficient. RAM requirements are exorbitant and update delays are obvious. A means to apply pre-filtering is necessary. A special data acquisition board is used with on-board RAM and in some cases on-board CPU. These boards range in price from 2000 to 5000 dollars. Up to 10,000 points per channel with up to 16 channels can be acquired on a board with 4Mb of RAM. This data must be filtered using an intelligent data communication driver. Keithley-Metrabyte's DAS-20 board is an example of such hardware.
The driver is a C-code program for communication between the board and RAM-resident datafiles. Functions to interpret shape and trend features in the data prepare symbols for the AI knowledge base. All functions and channels are configured using an ASCII initialization file. Heuristics ensure rapid computation of the required features and filtered data.13
CONCLUSION
AI techniques have become important computing tools to develop applications based on mental models of processes and domain knowledge. The efficiency of future methods will depend on the use of Computational Intelligence such as Fuzzy Logic, Artificial Neural networks and Genetic Algorithms. These methodologies form the foundation on which both biological and electronic intelligence depend.
Three methods to adapt Fuzzy Logic-based systems are available: automatic adjustment of a Fuzzy Set source using a collective variable, dynamic-switching of the Inference-Defuzzification method and adaptive Linguistic Defuzzification. Each technique provides elasticity to the output generated by a Fuzzy Logic Controller.
Artificial Neural Networks promise true learning capabilities but are unable on their own to explain the rational behind their conclusions. As modeling tools, they will likely play an important role in providing rapid pattern-matching capabilities. Holographic Networks bear examination as potential real-time learning tools.
Genetic Algorithms are one of the most efficient optimization techniques available to generate Fuzzy Set definitions or initial neural network weights.
Integration of these methods into Computationally Intelligent communication drivers will provide a means to provide accurate symbolic knowledge for processing by conventional AI knowledge bases in a supervisory fashion. Even data that must be sampled at an extremely rapid rate can be analysed by a conventional Expert System.
ACKNOWLEDGMENT
The author expresses his sincere appreciation to the students who participated in the industrial projects conducted at UBC over the past few years: Sunil Kumar, Lester Jordon, Paul Benford, Philippe Poirier, Marcello Veiga, Randy Gurton, Cliff Mui, Ken Scholey, Edgardo Cifuentes and Vladimir Rakocevic. Because of them, the Professor has become the Student. Support from Comdale Technologies, Toronto is also gratefully acknowledged.
REFERENCES
1. McDermott, K., Clyle, P., Hall, M. and Harris, C.A., 1992. "An Expert System for Control of No.4 Autogenous Mill Circuit at Wabush Mines", Proc. Canadian Mineral Processors, Paper 24, pp.20.
2. Freeman, N., Kemp, T. and Legg, J., 1990. "Development of Operator Guidance System for Pb-Blast Furnace Operations", Proc. Australia AI90 Conf., Perth, pp. 11.
3. Eggert, J., Folinsbee, J. and Benford, P., 1994. "SAG Mill Control at Dome Mines Using an Expert System", Proc. Canadian Mineral Processors, Ottawa, pp.16.
4. Benford, P. and Meech, J.A., 1992, "Advising Flotation Operators Using a Real-Time Expert System", Minerals Engineering, 5(10-12), 1325-1331.
5. Lacouture, B, Russell, C., Griffin, P. and Leung, K., 1991. "Copper Flotation Expert System at Mount Isa Mines Limited", Proc. Copper91, Eds. Dobby, G.S. , Argyropoulos and S.A., Rao, S.R. Conf. of Metallurgists, 429- 437.
6. Edwards, R. and Mular, A.L., 1992. "An Expert System Supervisor of a Flotation Circuit", CIM Bulletin, 84(959), 69-76.
7. Meech, J.A. and Kumar, S., 1994. "A HyperManual on Expert Systems - Version 3.0" CANMET Publication, Ottawa, 4723 electronic pages.
8. Harris, C.A. and Kosick, G.A. 1988. Expert System Technology at the Polaris Mine. Proc. Canadian Mineral Processors Conf., Ottawa, pp.25.
9. Poirier, P. and Meech, J.A., 1993. "Using Fuzzy Logic for On-Line Trend Analysis", Proc. 2nd IEEE Conference on Control Applications, Vancouver, Vol. 1, 83-86.
10. Poirier, P., Raabe, H. and Meech, J.A., 1993. Using Froth Identification in an Advisory Expert System for Copper Flotation Operations. Proc. Canadian Mineral Processors Conference, Ottawa, Paper 36, pp.14.
11. Meech, J.A. and Harris, C.A., 1992. "Expert Systems for Gold Processing Plants"; Randol Gold Forum, Vancouver, B.C., 45-61.
12. Harris, C.A. and Meech, J.A., 1987. Fuzzy Logic: A Potential Control Technique for Mineral Processing. CIM Bulletin., 80 (905), 51-59.
13. Rakocevic, V.M., Meech, J.A., Kumar, S., Gurton, R., Samarasekera, I.V. and Brimacombe, J.K., 1994. "Intelligent Computation in a SCADA System to Monitor the Continuous Casting of Steel Billets", Proc. CIM District 6 Meeting, Vancouver, 53-61.
14. Bezdek, J.C., 1994. "What is computational intelligence?", Computational Intelligence - Imitating Life., Proc. World Congress on Computational Intelligence, Orlando, Fla., IEEE Publications, 1-12.
15. Meech, J.A. and Jordon, L.A., 1993. "Development of a Self-Tuning Fuzzy Logic Controller", Minerals Engineering., 6(2), 119-131.
16. Smith, M.H., 1993. "Parallel Dynamic Switching of Reasoning Methods in a Fuzzy System", Proc. 2nd IEEE Inter. Conf. on Fuzzy Systems (FUZZ-IEEE'93), pp.968-973.
17. Smith, M.H. and Takagi, H., 1993. "Optimization of Fuzzy Systems by Switching Reasoning Methods Dynamically", Inter. Conf. on Fuzzy Systems, Seoul, Korea, pp.8.
18. Veiga, M.M. and Meech, J.A., 1994. "HgEx - A Heuristic System on Mercury Pollution in the Amazon." Water, Air & Soil Pollution, 14, 204-216.
19. Veiga, M.M., and Meech, J.A., 1994. "Application of Fuzzy Logic to Environmental Risk Assessment", Proc. IV Encuentro Hemisferio Sur sobre Tecnologia Mineral, Concepcion, Chile, 1994., pp.17.
20. Meech, J.A., Veiga, M.M and Hypolito, R., 1995. ""Educational Measures to Address Hg Pollution from Gold Mining Activities in the Amazon", AMBIOS (in press), pp. 15.
21. Veiga, M.M. and Meech, J.A., 1994. "Heuristic Approach to Mercury Pollution in the Amazon", Proc. International Sym. on Extraction and Processing for the Treatment and Minimization of Wastes, 123rd Congress of TMS, San Francisco, CA, 23-38
22. Zeidenberg, M., 1990. "Neural Networks in Artificial Intelligence", Ellis- Horwood Limited, Chichester, W. Sussex, England.
23. Touretzky, D.S. & Hinton, G.E., 1988. "A Distributed-Connectionist Production System", Cognitive Science, 12(3), 423 - 466.
24. Blum, A., 1992. "Neural Networks in C++", Wiley Professional Computing Series., New York, NY.
25. Rakocevic, V.M. and Meech, J.A., 1995. "An Artificial Neural Network to Interpret Froth Images from a Flotation Process", CIM Bulletin (in press), pp.16.
26. HNet Professional Users Manual, 1992. AND Corporation, Hamilton, Ontario, L8P 1C8
27. Soucek, B. (Ed.), 1993. "Fuzzy, Holographic and Parallel Intelligence", J. Wiley & Sons Inc., New York, NY.
28. Goldberg, D., 1989. "Genetic Algorithms in Search, Optimization & Machine Learning", Addison-Wesley, Reading, Mass., 1989.
29. Kumar, S., Meech, J.A., Samarasekera, I.V. and Brimacombe, J.K., 1993. "An Expert System to Trouble-Shoot Quality Problems in the Continuous Casting of Steel Billets", Iron & Steelmaker, ISS, 20(9), 29-36.
Return to the publications index