• No results found

CHAPTER 3 Maintenance Strategy and Operations and Its

4.3 Artificial Intelligence

The earliest pioneer in machine intelligence is Alan Turing, a mathematician and computer scientist, who queried whether the computer could think as human through his experiment called the Turing Test. He then wrote the most significant paper of “Computing Machinery and Intelligence” in 1950 (Negnevitsky, 2011). The term of Artificial Intelligence (AI) was introduced by John McCarthy in the 1950s in which Allen Newell, Herbert Simon, and J. C. Shaw introduced “Logic Theorist” which is recognised as the first AI program (Awad, 1996; Jones, 2008). Afterwards, the research on AI grew rapidly in many different fields.

4.3.1 From Intelligence to Artificial Intelligence

Intelligence is the ability to acquire and apply knowledge and skills (Oxford Dictionary, 2014). Jones (2008) defined it as a set of properties of the mind which enable it to make a right decision by being given a set of inputs and possible actions. When the human intelligence is imitated and their way of thinking is copied into a computer to enable it doing a cognitive task, it is called Artificial Intelligence (AI). Awad (1996) and Munakata (2008) similarly defined AI as the science of making computers do things that the human needs intelligence to do. But that is not the only way to define AI. Russell and Norvig (2010) summarised that AI can be defined from four different perspectives, acting humanly, thinking humanly, thinking rationally, and acting rationally.

The term of AI is commonly used in the computing science. The aim is to simulate the computer to work as human being to solve the problem. A simple hint to recognise AI is when a set of procedures enable a computer to see, hear, understand, and develop reasons to create a similar result as a human does, it can be called AI (Awad, 1996).

4.3.2 The Applications of Artificial Intelligence

The application of AI is widely ranging on different format and techniques to meet different requirement. Some of AI techniques which are widely used will be discussed in the following sections. These are Artificial Neural Network, Case-

P a g e | 46

Based Reasoning, Genetic Algorithm, Fuzzy System, and Expert System/Knowledge-Based System.

4.3.2.1 Artificial Neural Network

A neural network is a modelled system of processing elements of the human brain’s network called neurons (Awad, 1996). Since this method imitates the biological neuron process in the human brain, it is called Artificial Neural Network (ANN). ANN is intended to solve an interesting but cumbersome problem that cannot be carried on by traditional computer programme through mimicking networks as neurons and connecting them (Jones, 2008). The typical architecture of ANN is shown in Figure 4-3. The application of ANN has been attempted in the area of risk management, credit card fraud detection, check and signature verification, and mortgage appraisals (Awad, 1996).

Figure 4-3 Architecture of typical Artificial Neural Network (Negnevitsky, 2011) Each neuron is connected through links which have a numerical weight, also known as a long-term memory in ANN (Negnevitsky, 2011). Those neurons are arranged in the network along the layers. Input layers receive input signals from the outside and pass it to middle layers. The input signal could be in the form of raw data or output from other neurons. Middle layers then identify the features through the weight of input signals. The features are then translated into the pattern and sent to output layers. The output signals could be used as the final result or as the input for other neurons.

The transfer process, a so-called activation function, to weight the input signal, X, and compare it with a threshold value, θ, follows the equations below (Negnevitsky, 2011): In p u t si gn al s O u tp u t si gn al s

P a g e | 47

𝑋 = ∑𝑛𝑖=1 𝑥𝑖𝑤𝑖 Equation (1)

𝑌 = {+1 𝑖𝑓 𝑋 ≥ 𝜃

−1 𝑖𝑓 𝑋 < 𝜃 Equation (2)

where X is the net weighted input to the neuron, xi is the value of input i, wi is the

weight of input i, n is the number of the neuron inputs, and Y is the output of the neuron. This type of activation function is called a sign function. Thus the actual output of the neuron with a sign activation function can be represented as:

𝑌 = 𝑠𝑖𝑔𝑛[∑𝑛𝑖=1 𝑥𝑖𝑤𝑖 − 𝜃] Equation (3)

The ANN learning algorithms could be differentiated into two categories; supervised learning and unsupervised learning (Jones, 2008). Supervised learning algorithms train ANN to identify the right and wrong answers, find the error and adjust the weight, whereas the unsupervised learning algorithms do not guide the ANN to provide the specified answer (Mohamed, 2012). It is merely used to find the similarities and the differences among data and show the relationship of them.

4.3.2.2 Case-Based Reasoning

The essence of Case-Based Reasoning (CBR) is to recall the previous experience when the similar case arises (Hopgood, 2001). By this approach, CBR uses an existing database of old problem solution to address the problem on the new situation (Luger, 2009). This method is working properly for an expert who has experienced on a particular domain. Those experiences help him to reuse the reasoning of cases from the past to overcome the current problem (Awad, 1996). The solution sources to sort out the current problem could be generated from knowledge of engineering process or the previous result of problem-solving (Luger, 2009) by following a methodology as shown in Figure 4-4. CBR is widely used in fault diagnosis, engineering sales, help-desk support, and planning (Hopgood, 2001).

P a g e | 48

Figure 4-4 Methodology of Case-Based Reasoning, adapted from (Turban et al., 2007) and Luger (2009)

The old cases are indexed to help identify similarity of case features. The search engine tool in CBR system retrieves the appropriate cases heuristically from the case-base. It identifies the similar cases based on interpretation of both human and reasoning tool by referring to particular features. The CBR system then relates the old cases with the current problem and modifies those old cases based on current situation. The transformed case and modification solution are then applied to the current problem with a further iteration of “retrieve – modify – apply”. Finally, the implementation is documented and stored in the case-base, regardless of whether it a success or failure, as references for the future similar problems.

4.3.2.3 Genetic Algorithm

This type of AI is the population-based algorithm which is inspired by the biological process of evolution (Turban et al., 2007). Each cell of an individual organism contains chromosomes as the determinant of creature’s characteristics. A chromosome is made up of genes for which it determines a specific feature (Hopgood, 2001). The level of fitness of chromosome influences its chance to reproduce and pass its characteristics to the next generation.

In Genetic Algorithm (GA), chromosomes are known as the population of the candidate solution. Genes are represented as bit-strings of a binary number, one and zero (Negnevitsky, 2011). To find the solution space, GA uses the crossover, mutation, and inversion of evolutionary programming (Hopgood, 2001; Jones,

1. Assign indexes for old cases

2. Retrieve appropriate cases from memory

4. Apply the transformed case 5. Save the solution

for future use

3. Modify a retrieved case for the current

P a g e | 49

2008), as illustrated in Figure 4-5. Meanwhile, the basic steps of GA are shown in Figure 4-6.

Figure 4-5 Illustration of crossover, mutation, and inversion in GA (Jones, 2008)

Figure 4-6 Basic steps of Genetic Algorithm, adapted from Hopgood (2001) Initialisation of the population as a set of solution is generated randomly where an individual’s chromosome is evaluated based on the fitness function (Munakata, 2008). The fitter individual is more likely to be selected instead of the unfit ones. Reproduction of next generation is generated through crossover or recombination breeding between pairs of chromosomes and then applying mutation of the chromosome to each offspring. These iteration processes will be repeated until the best solution is achieved among them.

Furthermore, this method is followed by the development of Generic Programming and Evolutionary Strategies which replace the use of bit-strings as the genetic algorithm with other methods. Whilst Generic Programming uses S-

1111111111111111 0000000000000000 0000000011111111 1111111100000000 Single-Point Crossover Child B Child A Parent B Parent A 1111111111111111 0000000000000000 0000010000000000 1111111111110111 Mutation Child B Child A Parent B Parent A 1111111111111111 0000000000000000 1111000000000000 1111111100111111 Inversion Child B Child A Parent B Parent A Start Stop Randomly generate initial population Evaluate fitness of each

individual

Terminate? Select individuals to mate Generate offspring by

crossover Generate offspring by

mutation Replace old population

with new one

No

P a g e | 50

expressions or program trees as the encoding system, Evolutionary Strategies focuses on optimisation of real value parameters (Jones, 2008).

4.3.2.4 Fuzzy System

Fuzzy System (FS) deals with fuzzy logic which is related to the ambiguity, lack of definition and inexact knowledge with approximate rather than the exact logic of reasoning (Awad, 1996). It means that FS manages unclear boundaries on the particular problem by using the common-sense rules with uncertain data and infinite quantity. In a different way, Negnevitsky (2011) explained that FS represents the way of people think by modelling the sense of words, decision making, and common sense. Fuzzy logic uses the mathematical theory of fuzzy sets to allow the computer to behave less precisely and logically by considering not only right or wrong but also the “grey” area on the middle of them (Liao, 2005; Turban et al., 2007). For this reason, a fuzzy system is suitable to be applied when the conventional approach cannot provide the desired result and mathematical approach is hard to derive. It is usually used in the fuzzy control of physical and chemical characteristics, such as the motion of the machine, the flow of liquid or gas, and temperature (Munakata, 2008).

In the non-fuzzy set, the membership of element – called crisps, is only separated into two degrees, belongs to (1) and does not belong to the set (0) (Munakata, 2008), while in the fuzzy set, the crisps also contain some degrees between 1 and 0. The case example of “tall man” can be used to represent the fuzzy set. The crisps consist of three sets; short (150 – 170 cm), average (170 – 180 cm), and tall (> 180 cm), as illustrated in Figure 4-7a. If the boundary is set firmly, a 179 cm tall man is classified as the member of the average group, but a 181 cm tall man is classified as the member of the tall group. This distinction is not acceptable in common sense because this small difference of tall could significantly influence further treatment for him at the different group (Hopgood, 2001).

P a g e | 51

Figure 4-7 Illustration of crisp set (a) and fuzzy set (b) (Negnevitsky, 2011) By referring again to Figure 4-7, a 184cm tall man is not only to be classified as a member of the tall group. Instead, he is part of both average and tall groups but at different degree, as illustrated in Figure 4-7b. A 184 cm tall man is considered as a member of the average group with a degree of 0.1. At the same time, he is also part of a tall group with a degree of 0.6. By this, it means that he has partial membership in multiple sets (Negnevitsky, 2011).

4.3.2.5 Expert System/Knowledge-Based System

Expert System (ES) as an application of AI works intelligently as a human being in recommending advises of input problem (Liao, 2005). However, it is also known as part of Knowledge Management (KM) which involve all knowledge assets in its design (Becerra-Fernandez et al., 2004). ES is intended to act as an expert to consult a range problem in a particular and narrow area of expertise to create a solution (Hopgood, 2001). Therefore, the solution created should be complemented with the logic and reason to reach the conclusion. The conclusion produced is commonly called inference. By combining knowledge and inference, it is focusing on problem-solving by assisting the thinking process instead of merely providing information (Awad, 1996; Giarratano and Riley, 2005). When the knowledge is represented through generated and structured rules, inference represents the product of those rules to be gained by applying the ES.

The term of Expert System (ES) and Knowledge-Based (KB) System are used synonymously. On its development, the experts realised that even the most advanced expert system cannot beat the expert in dealing with given problems. Although ES could be powerful to gain a solution problem, it is only capable to face an individual case. To deal with another case, ES should be developed again from the beginning. The “brain” of ES is developed through the systematic

Short Average Tall

0.0 0.2 0.4 0.6 0.8 1.0 150 160 170 180 190 200 Height, cm D eg re e o f m em b er sh ip (a)

Short Average Tall

0.0 0.2 0.4 0.6 0.8 1.0 150 160 170 180 190 200 D eg re e o f m em b er sh ip (b)

P a g e | 52

algorithm of information and knowledge acquired by an expert. Considering that the Expert System is not truly capable of making a human expert as it is supposed to be, the term of ES is often replaced with Knowledge-Based System. Therefore in this research, the use of ES is presented as the KB System. As the method used in this research, the discussion about KB System in order to support the IMSO is detailed in the following sections.

Related documents