软件工程—外文翻译

软件工程—外文翻译 信息工程学院-软件工程 Artificial Immune Systems: A Novel Paradigm to Pattern Recognition Abstract This chapter introduces a new computational intelligence paradigm to perform pattern recognition, named Artificial Immune Systems (AIS). AIS take inspiration from the immune system in order to build novel computational tools to solve problems in a vast range of domain areas. The basic immune theories used to explain how the immune system perform pattern recognition are described and their corresponding computational models are presented. This is followed with a survey from the literature of AIS applied to pattern recognition. The chapter is concluded with a trade-off between AIS and artificial neural networks as pattern recognition paradigms. Keywords: Artificial Immune Systems;Negative Selection;Clonal Selection;Immune Network 1 Introduction The vertebrate immune system (IS) is one of the most intricate bodily systems and its complexity is sometimes compared to that of the brain. With the advances in the biology and molecular genetics, the comprehension of how the immune system behaves is increasing very rapidly. The knowledge about the IS functioning has unraveled several of its main operative mechanisms. These mechanisms have demonstrated to be very interesting not only from a biological standpoint, but also under a computational perspective. Similarly to the way the nervous system inspired the development of artificial neural networks (ANN), the immune system has now led to the emergence of artificial immune systems (AIS) as a novel computational intelligence paradigm. Artificial immune systems can be defined as abstract or metaphorical computational systems developed using ideas, theories, and components, extracted from the immune system. Most AIS aim at solving complex computational or engineering problems, such as pattern recognition, elimination, and optimization. This is a crucial distinction between AIS and theoretical immune system models. While the former is devoted primarily to computing, the latter is focused on the modeling of the IS in order to understand its behavior, so that contributions can be made to the biological sciences. It is not exclusive, however, the use of one approach into the other and, indeed, theoretical models of the IS have contributed to the development of AIS. This chapter is organized as follows. Section 2 describes relevant immune theories for pattern recognition and introduces their computational counterparts. In Section 3, we briefly describe how to model pattern recognition in artificial immune systems, and present a simple illustrative example. Section 4 contains a survey of AIS for pattern recognition, and Section 5 contrast the use of AIS with the use of ANN when applied to pattern recognition tasks. The chapter is concluded in Section 6. 2 Biological and Artificial Immune Systems 1 信息工程学院-软件工程 All living organisms are capable of presenting some type of defense against foreign attack. The evolution of species that resulted in the emergence of the vertebrates also led to the evolution of the immune system of this species. The vertebrate immune system is particularly interesting due to its several computational capabilities, as will be discussed throughout this section. The immune system of vertebrates is composed of a great variety of molecules, cells, and organs spread all over the body. There is no central organ controlling the functioning of the immune system, and there are several elements in transit and in different compartments performing complementary roles. The main task of the immune system is to survey the organism in the search for malfunctioning cells from their own body (e.g., cancer and tumour cells), and foreign disease causing elements (e.g., viruses and bacteria). Every element that can be recognized by the immune system is called an antigen (Ag). The cells that originally belong to our body and are harmless to its functioning are termed self (or self antigens), while the disease causing elements are named nonself (or nonself antigens). The immune system, thus, has to be capable of distinguishing between what is self from what is nonself; a process called self/nonself discrimination, and performed basically through pattern recognition events. From a pattern recognition perspective, the most appealing characteristic of the IS is the presence of receptor molecules, on the surface of immune cells, capable of recognising an almost limitless range of antigenic patterns. One can identify two major groups of immune cells, known as B-cells and T-cells. These two types of cells are rather similar, but differ with relation to how they recognise antigens and by their functional roles. B-cells are capable of recognising antigens free in solution (e.g., in the blood stream), while T-cells require antigens to be presented by other accessory cells. Antigenic recognition is the first pre-requisite for the immune system to be activated and to mount an immune response. The recognition has to satisfy some criteria. First, the cell receptor recognises an antigen with a certain affinity, and a binding between the receptor and the antigen occurs with strength proportional to this affinity. If the affinity is greater than a given threshold, named affinity threshold, then the immune system is activated. The nature of antigen, type of recognising cell, and the recognition site also influence the outcome of an encounter between an antigen and a cell receptor. The human immune system contains an organ called thymus that is located behind the breastbone, which performs a crucial role in the maturation of T-cells. After T-cells are generated, they migrate into the thymus where they mature. During this maturation, all T-cells that recognise self-antigens are excluded from the population of T-cells; a process termed negative selection. If a B-cell encounters a nonself antigen with a sufficient affinity, it proliferates and differentiates into memory and effector cells; a process named clonal selection. In contrast, if a B-cell recognises a self-antigen, it might result in suppression, as proposed by the immune network theory. In the following subsections, each of these processes (negative selection, clonal selection, and network theory) will be described separately, along with their computational algorithms counterparts. 2.1 Negative Selection 2 信息工程学院-软件工程 The thymus is responsible for the maturation of T-cells; and is protected by a blood barrier capable of efficiently excluding nonself antigens from the thymic environment. Thus, most elements found within the thymus are representative of self instead of nonself. As an outcome, the T-cells containing receptors capable of recognising these self antigens presented in the thymus are eliminated from the repertoire of T-cells through a process named negative selection. All T-cells that leave the thymus to circulate throughout the body are said to be tolerant to self, i.e., they do not respond to self. From an information processing perspective, negative selection presents an alternative paradigm to perform pattern recognition by storing information about the complement set (nonself) of the patterns to be recognised (self). A negative selection algorithm has been proposed in the literature with applications focused on the problem of anomaly detection, such as computer and network intrusion detection, time series prediction, image inspection and segmentation, and hardware fault tolerance. Given an appropriate problem representation (Section 3), define the set of patterns to be protected and call it the self- set (P). Based upon the negative selection algorithm, generate a set of detectors (M) that will be responsible to identify all elements that do not belong to the self-set, i.e., the nonself elements. After generating the set of detectors (M), the next stage of the algorithm consists in monitoring the system for the presence of nonself patterns (Fig 2(b)). In this case, assume a set P* of patterns to be protected. This set might be composed of the set P plus other new patterns, or it can be a completely novel set. For all elements of the detector set, that corresponds to the nonself patterns, check if it recognises (matches) an element of P* and, if yes, then a nonself pattern was recognized and an action has to be taken. The resulting action of detecting nonself varies according to the problem under evaluation and extrapolates the pattern recognition scope of this chapter. 2.2 Clonal Selection Complementary to the role of negative selection, clonal selection is the theory used to explain how an immune response is mounted when a nonself antigenic pattern is recognised by a B-cell. In brief, when a B-cell receptor recognises a nonself antigen with a certain affinity, it is selected to proliferate and produce antibodies in high volumes. The antibodies are soluble forms of the B-cell receptors that are released from the B-cell surface to cope with the invading nonself antigen. Antibodies bind to antigens leading to their eventual elimination by other immune cells. Proliferation in the case of immune cells is asexual, a mitotic process; the cells divide themselves (there is no crossover). During reproduction, the B-cell progenies (clones) undergo a hyper mutation process that, together with a strong selective pressure, result in B-cells with antigenic receptors presenting higher affinities with the selective antigen. This whole process of mutation and selection is known as the maturation of the immune response and is analogous to the natural selection of species. In addition to differentiating into antibody producing cells, the activated Bcells with high antigenic affinities are selected to become memory cells with long life spans. These memory cells are pre-eminent in future responses to this same antigenic pattern, or a similar one. 3 信息工程学院-软件工程 Other important features of clonal selection relevant from the viewpoint of computation are: 1. An antigen selects several immune cells to proliferate. The proliferation rate of each immune cell is proportional to its affinity with the selective antigen: the higher the affinity, the higher the number of offspring generated, and vice-versa; 2. In complete opposition to the proliferation rate, the mutation suffered by each immune cell during reproduction is inversely proportional to the affinity of the cell receptor with the antigen: the higher the affinity, the smaller the mutation, and vice-versa. Some authors have argued that a genetic algorithm without crossover is a reasonable model of clonal selection. However, the standard genetic algorithm does not account for important properties such as affinity proportional reproduction and mutation. Other authors proposed a clonal selection algorithm, named CLONALG, to fulfil these basic processes involved in clonal selection. This algorithm was initially proposed to perform pattern recognition and then adapted to solve multi-modal optimisation tasks. Given a set of patterns to be recognised (P), the basic steps of the CLONALG algorithm are as follows: 1. Randomly initialise a population of individuals (M); 2. For each pattern of P, present it to the population M and determine its affinity (match) with each element of the population M; 3. Select n1 of the best highest affinity elements of M and generate copies of these individuals proportionally to their affinity with the antigen. The higher the affinity, the higher the number of copies, and vice-versa; 4. Mutate all these copies with a rate proportional to their affinity with the input pattern: the higher the affinity, the smaller the mutation rate, and vice-versa. 5. Add these mutated individuals to the population M and reselect n2 of these maturated (optimised) individuals to be kept as memories of the system; 6. Repeat Steps 2 to 5 until a certain criterion is met, such as a minimum pattern recognition or classification error. Note that this algorithm allows the artificial immune system to become increasingly better at its task of recognising patterns (antigens). Thus, based upon an evolutionary like behaviour, CLONALG learns to recognise patterns. 2.3 Immune Network The immune network theory proposes that the immune system has a dynamic behaviour even in the absence of external stimuli. It is suggested that the immune cells and molecules are capable of recognising each other, what endows the system with an eigen behaviour that is not dependent on foreign stimulation. Several immunologists have refuted this theory, however its computational aspects are relevant and it has proved itself to be a powerful model for computational systems. According to the immune network theory, the receptor molecules contained in the surface of the immune cells present markers, named idiotopes, which can be recognized by receptors on other immune cells. These idiotopes are displayed in and/or around the same portions of the receptors that recognise nonself antigens. To explain the network theory, assume that a receptor (antibody) Ab1 on 4 信息工程学院-软件工程 a B-cell recognises a nonself antigen Ag. Assume now, that this same receptor Ab1 also recognises an idiotope i2 on another B-cell receptor Ab2. Keeping track of the fact that i2 is part of Ab2, Ab1 is capable of recognising both Ag and Ab2. Thus, Ab2 is said to be the internal image of Ag, more precisely, i2 is the internal image of Ag. The recognition of idiotopes on a cell receptor by other cell receptors, lead to ever increasing sets of connected cell receptors and molecules. Note that the network in this case, is a network of affinities, which different from the ‘hardwired’ network of the nervous system. As a result of the network recognition events, it was suggested that the recognition of a cell receptor by another cell receptor results in network suppression, whilst the recognition of an antigen by a cell receptor results in network activation and cell proliferation. The original theory did not account explicitly for the results of network activation and/or suppression, and the various artificial immune networks found in the literature model it in a particular form. 3 Modelling Pattern Recognition in AIS Up to this point, the most relevant immune principles and their corresponding computational counterparts to perform pattern recognition have been presented. In order to apply these algorithms to computational problems, there is a need to specify a limited number of other aspects of artificial immune systems, not as yet covered. The first aspect to introduce is the most relevant representations to be applied to model self and nonself patterns. Here the self-patterns correspond to the components of the AIS responsible for recognising the input patterns (nonself). Secondly, the mechanism by which the evaluation of the degree of match (affinity), or degree of recognition, of an input pattern by an element of the AIS has to be discussed. To model immune cells, molecules, and the antigenic patterns, the shape-space approach proposed is usually adopted. Although AIS model recognition through pattern matching, given certain affinity functions to be described further, performing pattern recognition through complementarity or similarity is based more on practical aspects than on biological plausibility. The shape-space approach proposes that an attribute string s = ás1, s2,…,sLñ in an Ldimensional shape-space, S, (s Î SL), can represent any immune cell or molecule. Each attribute of this string is supposed to represent a feature of the immune cell or molecule, such as its charge, van der Wall interactions, etc. In the development of AIS the mapping from the attributes to their biological counterparts is usually not relevant. The type of attributes used to represent the string will define partially the shape-space under study, and is highly dependent on the problem domain. Any shape-space constructed from a finite alphabet of length k constitutes a k-ary Hamming shape-space. As an example, an attribute string built upon the set of binary elements {0,1} corresponds to a binary Hamming shape-space. It can be thought of, in this case, of a problem of recognising a set of characters represented by matrices composed of 0’s and 1’s. Each element of a matrix corresponds to a pixel in the character. If the elements of s are represented by real-valued vectors, then we have an Euclidean shape-space. Most of the AIS found in the literature employ binary Hamming or Euclidean shape-spaces. Other types of shape-spaces are also possible, such as symbolic shape-spaces, which combine different (symbolic) attributes in the representation of a single string s. These are usually found in data mining applications, where the data might contain symbolic information like age, name, etc., of a set of patterns. 5 信息工程学院-软件工程 Another important characteristic of the artificial immune systems is that most of them are population based. It means that they are composed of a set of individuals, representing immune cells and molecules, which have to perform a given role; in our context, pattern recognition. If we recapitulate the three immune processes reviewed, negative selection, clonal selection, and immune network, all of them rely on a population M of individuals to recognise a set P of patterns. The negative selection algorithm has to define a set of detectors for nonself patterns; clonal selection reproduces, maturates, and selects self-cells to recognise a set of nonself; and the immune network maintains a set of individuals, connected as a network, to recognize self and nonself. Consider first the binary Hamming shape-space case, which is the most widely used. There are several expressions that can be employed in the determination of the degree of match or affinity between an element of P and an element of M. The simplest case is to simply calculate the Hamming distance (DH) between these two elements, as given by Eq. (1). Another approach is to search for a sequence of r-contiguous bits, and if the number of r-contiguous matches between the strings is greater than a given threshold, then recognition is said to have occurred. As the last approach to be mentioned here, we can describe the affinity measure of Hunt, given by Eq. (2). This last method has the advantage that it favours sequences of complementary matches, thus searching for similar regions between the attribute strings (patterns). L1ifpm,,iiDwhere,,, (1) ,,,,H0otherwise,1i, li (2) DD,，2,Hi lwhere is the length of the i-th sequence of matching bits longer than 2. i In the case of Euclidean shape-spaces, the Euclidean distance can be used to evaluate the affinity between any two components of the system. Other approaches such as the Manhattan distance may also be employed. Note that all the methods described rely basically, on determining the match between strings. However, there are AIS in the literature that take into account other aspects, such as the number of patterns matched by each antibody. 4 A Survey of AIS for Pattern Recognition The applications of artificial immune systems are vast, ranging from machine learning to robotic autonomous navigation. This section will review some of the works from the AIS literature applied to the pattern recognition domain. The rationale is to provide a guide to the literature and a brief description of the scope of applications of the algorithms. The section is divided into two parts for ease of comprehension: 1) computer security, and 2) other applications. The problem of protecting computers (or networks of computers) from viruses, unauthorised users, etc., constitutes a rich field of research for pattern recognition systems. Due, mainly, to the appealing intuitive metaphor of building artificial immune systems to detect computer viruses, there has been a great interest from the computer science community to this particular application. The use of the negative and clonal selection algorithms have been widely tested on this application. The former because it is 6 信息工程学院-软件工程 an inherent anomaly (change) detection system, constituting a particular case of a pattern recognition device. The latter, the clonal selection algorithm, has been used in conjunction to negative selection due to its learning capabilities. Other more classical pattern recognition tasks, such as character recognition, and data analysis have also been studied within artificial immune systems. 5 AIS and ANN for Pattern Recognition Similar to the use of artificial neural networks, performing pattern recognition with an AIS usually involves three stages: 1) defining a representation for the patterns; 2) adapting (learning or evolving) the system to identify a set of typical data; and 3) applying the system to recognise a set of new patterns (that might contain patterns used in the adaptive phase). Refering to the three immune algorithms presented (negative selection, clonal selection, and immune network), coupled with the process of modelling pattern recognition in the immune system, as described in Section 3, this section will contrast AIS and ANN focusing the pattern recognition applications. Discussion will be based on computational aspects, such as basic components, adaptation mechanisms, etc. Common neural networks for pattern recognition will be considered, such as single and multi-layer perceptrons, associative memories, and self-organising networks. All these networks are characterised by set(s) of units (artificial neurons); they adapt to the environment through a learning (or storage) algorithm, they can have their architectures dynamically adapted along with the weights, and they have the basic knowledge stored in the connection strengths. Component: The basic unit of an AIS is an attribute string s (along with its connections in network models) represented in the appropriate shape-space. This string s might correspond to an immune cell or molecule. In an ANN, the basic unit is an artificial neuron composed of an activation function, a summing junction, connection strengths, and an activation threshold. While artificial neurons are usually processing elements, attribute strings representing immune cells and molecules are information storage and processing components. Location of the components: In immune network models, the cells and molecules usually present a dynamic behaviour that tries to mimic or counteract the environment. This way, the network elements will be located according to the environmental stimuli. Unlike the immune network models, ANN have their neurons positioned in fixed predefined locations in the network. Some neural network models also adopt fixed neighbourhood patterns for the neurons. If a network pattern of connectivity is not adopted for the AIS, each individual element will have a position in the population that might vary dynamically. Also, a metadynamic process might allow the introduction and/or elimination of particular units. Structure: In negative and clonal AIS, the components are usually structured around matrices representing repertoires or populations of individuals. These matrices might have fixed or variable dimensions. In artificial immune networks and artificial neural networks, the components of the population are interconnected and structured around patterns of connectivity. Artificial immune networks usually have an architecture that follows the spatial distribution of the antigens represented in shape-space, while ANN usually have pre-defined architectures, and weights biased 7 信息工程学院-软件工程 by the environment. Memory: The attribute strings representing the repertoire(s) of immune cells and molecules, and their respective numbers, constitute most of the knowledge contained in an artificial immune system. Furthermore, parameters like the affinity threshold can also be considered part of the memory of an AIS. In artificial immune network models, the connection strengths among units also carry endogenous and exogenous information, i.e., they quantify the interactions of the elements of the AIS themselves and also with the environment. In most cases, memory is content-addressable and distributed. In the standard (earliest) neural network models, knowledge was stored only in the connection strengths of individual neurons. In more sophisticate strategies, such as constructive and pruning algorithms, and networks with self-adaptive parameters, the final number of network layers, neurons, connections, and the shapes of their respective activation functions are also part of the network knowledge. The memory is usually self-associative or content-addressable, and distributed. Adaptation: Adaptation usually refers to the alteration or adjustment in the structure or behaviour of a system so that its pattern of response to other components of the system and to the environment changes. Although both evolutionary and learning processes involve adaptation, there is a conceptual difference between them. Evolution can be seen as a change in the genetic composition of a population of individuals during successive generations. It is a result of natural selection acting on the genetic variation among individuals. In contrast, learning can be seen as a long lasting change in behaviour as a result of previous experience. While AIS might present both types of adaptation, learning and evolution, ANNs adapt basically through learning procedures. Plasticity and diversity: Metadynamics refers basically to two processes: 1) the recruitment of new components into the system, and 2) the elimination of useless elements from the system. As consequences of metadynamics, the architecture of the system can be more appropriately adapted to the environment, and its search capability (diversity) increased. In addition, metadynamics reduces redundancy within the system by eliminating useless components. Metadynamics in the immune algorithms corresponds to a continuous insertion and elimination of the basic elements (cells/molecules) composing the system. In ANN, metadynamics is equivalent to the pruning and/or insertion of new connections, units, and layers in the network. Interaction with other components: The interaction among cells and molecules in AIS occurs through the recognition (matching) of attribute strings by cell receptors (other attribute strings). In immune network models, the cells usually have weighted connections that allow them to interact with (recognise and be recognised by) other cells. These weights can be stimulatory or suppressive indicating the degree of interaction with other cells. Artificial neural networks are composed of a set (or sets) of interconnected neurons whose connection strengths assume any positive or negative values, indicating an excitatory or inhibitory activation. The interaction with other neurons in the network occurs explicitly through these connection strengths, where a single neuron receives and processes inputs from the environment (or network neurons) in the same or other layer(s). An individual neuron can also receive an input from itself. Interaction with the environment: In pattern recognition applications, the environment is sually represented as a set of input patterns to be learnt, recognised, and/or classified. In AIS, an attribute 8 信息工程学院-软件工程 string represents the genetic information of the immune cells and molecules. This string is compared with the patterns received from the environment. If there is an explicit antigenic population to be recognised (set of patterns), all or some antigens can be presented to the whole or parts of the AIS. At the end of the learning or recognition phase, each component of the AIS might recognise some of the input patterns. The artificial neurons have connections that receive input signals from the environment. These signals are processed by neurons and compared with the information contained in the artificial neural network, such as the connection strengths. After learning, the whole ANN might (approximately) recognise the input patterns. Threshold: Under the shape-space formalism, each component of the AIS interacts with other cells or molecules whose complements lie within a small surrounding region, characterised by a parameter named affinity threshold. This threshold determines the degree of recognition between the immune cells and the presented input pattern. Most current models of neurons include a bias (or threshold). This threshold determines the neuron activation, i.e., it indicates how sensitive the neuron activation will be with relation to the input signal. Robustness: Both paradigms are highly robust due mainly to the presence of populations or networks of components. These elements, cells, molecules, and neurons, can act collectively, co-operatively, and competitively to accomplish their particular tasks. As knowledge is distributed over the many components of the system, damage or failure to individual elements might not significantly deteriorate the overall performance. Both AIS and ANN are highly flexible and noise tolerant. An interesting property of immune network models and negative selection algorithms is that they are also self-tolerant, i.e., they learn to recognise themselves. In immune network models, the cells interact with each other and usually present connection strengths quantifying these interactions. In negative selection algorithms, the self-knowledge is performed by storing information about its complement. State: At each iteration, time step or interval, the state of an AIS corresponds to the concentration of the immune cells and molecules, and/or their affinities. In the case of immune network models, the connection strengths among units are also part of the current state of the system. In artificial neural networks, the activation level of the output neurons determines the state of the system. Notice that this activation level of the output neurons takes into account the number of connection strengths and their respective values, the shape of activation functions and the network dimension. Control: Any immune principle, theory or process can be used to control the types of interaction among the many components of an AIS. As examples, clonal selection can be employed to build an antibody repertoire capable of recognising a set of antigenic patterns, and negative selection can be used to define a set of antibodies (detectors) for the recognition of anomalous patterns. Differential or difference equations can be applied to the control of how an artificial immune network will interact with itself and the environment. Basically, three learning paradigms can be used to train an ANN: 1) supervised, 2) unsupervised, and 3) reinforcement learning. Generalisation capability: In the AIS case, cells and molecules capable of recognising a certain pattern, can recognise not only this specific pattern, but also any structurally related pattern. 9 信息工程学院-软件工程 This capability is attained by a process called cross-reactivity, and can be modelled using the affinity threshold. Any pattern lying in a ‘neighbourhood’ of a known pattern can be recognised by the same component of the AIS that recognise the known pattern. Thus, a component of the AIS can generally recognise any other element whose affinity with is superior to e. In addition to cross-reactivity, some immunologists, speculate that antibodies can also be multi-specific, in the sense that they can recognise antigens of relatively different structures, as far as enough interactions are established between them. Therefore, multispecificity contributes to the generalization capability of AIS. ANNs are known to be efficient in generalising the training patterns, provided that an appropriate learning is performed. There are basically two ways in which an ANN can attain a satisfactory generalisation performance: 1) by reducing the number of dimensions of the parameter space, or 2) by reducing the effective size of each dimension. Non-linearities: Non-linearities in AIS appear basically in the use of activation functions that define the degree of recognition between two components of the system, proportionally to their affinity. As examples, a sigmoid or a simple threshold matching function might be used. Some immune network models use Gaussian-like functions to make the maturation and proliferation probabilities dependent on the degree of connectivity of an immune cell with the current network configuration. Non-linearities in artificial neural networks reside basically in the activation functions of individual neurons. The ensemble operation of several non-linear neural units results in a network with great potentials to perform non-linear approximations and/or classifications. 6 Concluding Remarks Artificial immune systems constitute an emergent biologically motivated computing paradigm. It is based upon the extraction of principles and metaphors from the immune system in order to design alternative computational tools to solve complex problems. Indeed, the main role of the immune system is to recognise what cells, molecules, and tissues belong to the organism and to distinguish them from the foreign elements. If the immune system were not so efficient in this self/nonself discrimination process, the body would have no problem with the rejection of graft tissues, for example. As a consequence, this great capability to recognise and eliminate specific patterns (nonself) serves as a good source of inspiration to develop novel computational paradigms for machine-learning and pattern recognition. In this chapter three classes of artificial immune system algorithms to perform pattern recognition: 1) negative selection, 2) clonal selection, and 3) immune network models, have been reviewed. In negative selection, a pattern recognition system is designed by learning information about the complement set of the patterns to be recognised - a brand new paradigm. Clonal selection algorithms learn to recognize patterns through an evolutionary-like procedure. Finally, immune network models are peculiar because they carry information about the patterns to be recognised and, also, they have knowledge of themselves, i.e., a notion of self-identification. All algorithms are population based with the knowledge distributed among the components of the system. The intuitive and appealing metaphor of engineering artificial immune systems to protect computers and networks of computers from viruses, unauthorised users, etc., led to the development 10 信息工程学院-软件工程 of the so-called computational immunology. Most computational immunology algorithms, which compose particular cases of artificial immune systems, are based upon the negative selection algorithm. In the survey section of this chapter, the most influential works in computational immunology we reviewed. Additionally, the application of other models, including the immune network and clonal selection algorithms, to other types of pattern recognition applications, such as character recognition, data analysis, clustering and classification were discussed. The chapter followed with a theoretical comparison between artificial immune systems and neural network models for pattern recognition. Aspects such as the basic units composing each system, their respective types of adaptation mechanisms, the types of memory presented, and how they present generalisation capabilities were stressed. There are also several works in the literature hybridising neural networks with artificial immune systems; Although these were not included here due to a lack of space, these authors strongly believe that both approaches have much to profit from one another. Stretching speculations, it could be suggested that novel paradigms will soon emerge, such as artificial neuroimmune systems. The aim of this chapter was to serve the purpose of introducing artificial immune systems to the neural network community, and also provided a basic guide to the literature. The algorithms presented could be directly employed and/or adapted as alternatives to solve the same types of pattern recognition problems as neural networks, or to complement their potentialities. 11 信息工程学院-软件工程 人工免疫系统:一种范式模式识别方法 摘 要 本章所介绍的是一个新的智能计算范式---一种可以进行识别的范式,称之为人工免疫系统(AIS)。人工免疫系统采取的灵感来自免疫系统，目的是为了在广阔的领域中建立一个新型的计算工具去解决问题。基本的免疫系统原理经常被用来理解如何使用免疫系统的执行模式识别进行了描述,并给出了相应的计算模型。这是通过从文学上对AIS的调查来识别。本文是作为一个权衡AIS和人工神经网络的模式来进行识别的范例。 关键字:人工免疫系统;否定选择;克隆选择;免疫网络 1 引言 脊椎动物的免疫系统是其中最复杂的身体系统，其复杂性有时可以和人类的大脑相比。生物学和分子遗传学的进展对如何提高免疫系统的行为是非常有力的。脊椎动物的免疫系统的知识作用有它的一些主要的有效机制。不管是从生物学角度看,还是从计算的观点来看，这些机制都被证明是非常有趣的。同时对神经系统的发展与人工神经网络(ANN),免疫系统作为一种计算智能的范例引起了人工免疫系统的出现(AIS)。 人工免疫系统可以被定义为一种抽象或隐喻性的计算系统开发，它是利用了想法、理论、成分、提取的免疫系统。大多数AIS旨在解决复杂的计算和工程问题,如模式识别、排除、和优化。这是一个非常重要的免疫系统的系统模型。前者是后者的计算,主要集中在建模，是为了了解它的行为,所以它的有用性可以使用在生物科学的领域。它不是专属的,然而,却是常用的使用方法之一,而且确实对免疫系统的发展的理论模型是有贡献的。 这一章组织如下;第二节描述了对模式识别和免疫系统的介绍了以及他们相关理论的计算。在第三节,我们简要描述了如何在人工免疫系统进行模式识别,提出了一种简单的数值例子。第四节包含了一个调查,对模式识别、免疫系统的使用,第5节对比AIS和ANN用在模式识别的任务。第6部分是。 2生物免疫系统与人工免疫系统 所有的生物都能呈现出某种类型的防御来阻止外来的攻击。物种的进化导致了脊椎动物免疫系统的演化。脊椎动物的免疫系统是非常复杂的,原因是它的计算能力,这种能力将在这部分进行讨论。 脊椎动物的免疫系统是由各种各样的分子、细胞和器官组成的,遍及整个身体。没有中央机构的控制作用的免疫系统,并有几种元素在随时发生改变、在各分区执行互补的作用。免疫系统最主要的任务是生物体在自己的身体寻找故障的细胞(如癌症和肿瘤细胞),与身体之外的致病因素(例如,细菌和病毒)。每个细胞都可以被免疫系统称为抗原 (Ag)。这些细胞,原来属于我们的身体,是无害的,它的功能是自我防御(或自我抗原),而致病因素被命名为异物(或非抗原)。因此,免疫系统必须有能够区分什么是自己什么事外来的能力.这个过程叫做自我识别,并能够完成对主要事件的模式识别。 从一个模式识别的角度来看,最吸引人的特点是存在着的受体分子,面上的免疫细胞,能够识别几乎无限的抗原的模式。你可以看到两大类,已知的B免疫细胞和T免疫细胞。这两 12 信息工程学院-软件工程 种类型的细胞是颇为相似的,但是它们的不同与他们如何识别抗原和他们的职能作用有关。B细胞能够在人体内自由识别抗原,当T细胞需要抗原时,会由其他辅助细胞产生抗原。 抗原识别是第一前，必须通过必要的免疫系统被激活并安装的免疫反应。承认必须满足一些条件。首先，细胞受体识别具有一定的亲和力的抗原，而与该受体与抗原结合的强度成正比，出现这种亲和力。如果亲和力是一个给定的阈值比更大，名为亲和力阈值，那么免疫系统被激活。抗原的性质，认识到细胞类型，识别抗原与细胞受体。 人类的免疫系统包含一个名为器官是背后的胸骨，它执行一个在T细胞成熟的关键作用所在胸腺。后T细胞产生，他们迁移到胸腺他们成熟。在这个成熟，所有的T -细胞识别自身抗原被排除在T细胞的人口;一个过程称为负选择。如果B细胞遇到了足够的亲和力1非我抗原，它的扩散并记忆和效应细胞的区别;一个过程称为克隆选择。相反，如果B细胞能识别自身抗原，它可能会导致抑制，经免疫网络理论提出。在下面的小节中，这些进程(负选择，克隆选择，网络理论)将独立出来介绍，以及它们对应的计算算法。 2.1 否定选择 胸腺是为T细胞成熟负责，以及由血液屏障有效地排除非己抗原的胸腺能够保护环境。因此，大多数元素的胸腺内发现的代表，而不是自我非我的。作为一个结果，T细胞受体识别包含在胸腺能够提出这些自我抗原的剧目是从淘汰的T通过命名细胞阴性选择的过程。所有的T -细胞离开胸腺整个身体散发据说可以容忍自己，也就是说，它们不应对自己。 从信息处理的角度，提出了一种消极选择的另一种模式来执行约的补集(非我的模式)存储信息模式识别予以承认(自我)。负选择算法已经在与有关的异常检测，如计算机和网络入侵检测，时间序列预测，图像检测与分割，文献中提出的问题，侧重于应用和硬件容错功能。给定一个适当的代表权问题，定义的模式设置为保护，并调用它的自我集(P)。基于否定选择算法，生成一个探测器集，将负责确定所有元素不属于自己设定的，即非我元素。 生成后的探测器集，该算法在下一阶段包括监测为非我模式。在这种情况下，假设一个集合P的模式，以保障P*。这个集合可能组成的集合P加上其它新花样，也可以是一个完全新的一套。 对于探测器集的所有元素，对应于非我模式，检查它是否承认(匹配)P的元素P*，如果是，那么非我模式是公认的和将要采取行动。该检测产生的行动非我根据不同情况进行评价的问题，并推断本章模式识别的范围。 2.2 克隆选择 为了配合角色的负选择，克隆选择是用来解释如何安装一个免疫反应的抗原时，非我模式是由B细胞所承认的理论。简言之，当B细胞受体识别具有一定的亲和力的非我抗原，它被选中的增殖并产生高量的抗体。该抗体的B细胞是从B -细胞表面释放，以应付入侵的抗原受体非我可溶性形式。抗体结合抗原领先其他免疫细胞最终将其消灭。在免疫细胞的增殖是无性的情况，有分裂过程;自己的细胞分裂(没有交叉)。在复制，B细胞后代(克隆)接受一个超突变的过程中，与一个强大的选择压力，在B -细胞的结果与抗原呈现较高的选择性亲和抗原受体。这种突变的整个过程和选择被看作是成熟的免疫反应称为[35]和类似于物种的自 13 信息工程学院-软件工程 然选择。除了为抗体生成细胞分化，具有高亲和力的抗原激活B细胞被选中成为长寿命的记忆细胞。这些存储单元是预先在此相同的抗原模式，或类似的一个未来的反应突出。 其他具有克隆选择的相关重要特征的计算包括: 1.选择一个抗原的免疫细胞增殖数。在每个免疫细胞增殖率是成正比的亲和力和选择性抗原:高亲和力,更高的子代的数量,反之亦然。 2. 在完全反对核扩散的突变速率,每个免疫细胞中遭受的复制成反比的细胞受体的亲和力的抗原:高亲和力,规模较小的突变,反之亦然。 一些作者人认为，无交叉遗传算法是一种克隆选择合理的模型。然而，遗传算法不占比例，如亲和繁殖和变异的重要特性。提出了克隆选择算法，命名为克隆选择，克隆选择，以履行其在参与这些基本过程的其他作者。该算法最初提出来执行模式识别并适用于解决多模态优化任务。由于一组模式予以承认的，在克隆选择算法的基本步骤如下:: 1. 随机生成一个初始抗体集; 2. 对于每一个模式的P,和人口M和确定它的亲和力,(比较)每个元素的亲和度; 3. 选择n1最好的最高亲和力的元素,产生这些个体对他们结亲比例的抗原。更高的亲 和力,较高的份数,反之亦然。 4. 所有这些拷贝和变异率成正比的亲和力和输入模式:高亲和力,规模较小的突变速率, 反之亦然。 5. 把这些突变的抗体数M和重新添加到第n2期的优秀的(优化)抗体可以作为记忆的 系统。 6. 重覆执行步骤2至5至一定标准,如一个最小的模式识别和分类的错误。 注意这个算法允许人工免疫系统,在它的任务变得越来越更好的识别模式(抗原)。因此,基于进化像行为,克隆学会识别方式。 2.3 免疫网络 免疫网络理论认为人体的免疫系统，甚至在没有外来刺激的动态行为。这是建议，免疫细胞和分子识别对方的能力，赋予了自我识别是不依赖于外部刺激系统。几个免疫学家驳斥这种理论，但其计算方面具有相关性，它已证明自己是一个强大的计算系统模型。 根据免疫网络理论,提出了一种受体分子包含在表面的免疫细胞,进行独特的目前标记,它可以被其他免疫细胞上的受体结合。这些独特的陈列会在同一份的受体承认非抗原。网络理论来解释,假定一个受体(抗体)Ab1在体液抗原。现在,这同样假定一个独特的受体Ab1同样可以在另一个体液受体Ab2感应。跟踪这样的事实Ab2感应的一部分,Ab1能够识别两银和Ab2。因此,Ab2据说是内部图像的银,更准确地说,是内部图像的感应。独特的识别在细胞受体的其他细胞的受体,导致日益套连接的细胞受体和分子。需要注意的是,在这种情况下,网络是一个网络的亲缘关系,这不同于“天生的网络的神经系统。由于网络识别事件,这就暗示了识别的一个细胞受体被另一个细胞受体的结果,同时在网络抑制的认定,由细胞受体的结果抗原在细胞增殖及网络激活。原有的理论不能明确的结果网络激活和/或抑制、各种人工免疫网络文献中 14 信息工程学院-软件工程 找到它,在某一特定形式的模型。 3人工免疫系统的模式识别 至此，最相关的免疫原则及其相应的计算同行进行模式识别已提交。为了应用这些算法的计算问题，有必要指定的其他方面的人工免疫系统，而不是少数尚未包括在内。第一个方面是引进最相关的模型被应用到自我和非我的模式交涉。在这里，自我模式对应的客户资料安全负责确认输入模式(非我)负责的组成部分。其次，该机制的匹配度(亲和力评估)，或认可程度，由一个输入模式对认可机构的元素，需要讨论。为了模型的免疫细胞，分子及抗原图案，形状，空间的办法通常是采用建议。虽然认可机构的承认，通过模式匹配模型，给予一定的亲和职能得到进一步的叙述，通过互补性或相似性表演模式识别是基于生物可信度比实际方面的问题。形状空间的做法提出了一个属性字符串s = ás1和S2，...，在一有限形状的空间，S前哨，可以代表任何免疫细胞或分子。此字符串的每个属性应该是代表着一种免疫细胞或分子，如它的收费，特点，范德壁相互作用，是等在AIS的发展从属性映射到对应的生物，通常不相关的。属性的类型用来表示字符串将确定部分的形状，正在研究空间，是高度依赖的问题域。任何形状的空间从一个长度为k有限字母表兴建构成k元海明形状的空间。作为一个例子，一个属性字符串建基于二进制元素的集合(0,1)对应的二进制海明形状的空间。它可以被认为，在这种情况下，是承认一个由0,1组成的矩阵表示的字符集和1的问题。矩阵的每一个元素对应一个像素中的字符。如果S的元素是由实值向量代表，那么我们有一个形状欧几里得空间。大部分认可机构在文献中找到二进制海明或聘用欧几里德形状空间。形状，空间的其他类型也是可能的，例如象征性形状的空间，这结合不同(符号)在一个字符串s的代表属性这些通常是发现在数据挖掘应用程序，其中可能包含的数据符号信息，如年龄，姓名等的一组模式。 另一种人工免疫系统的一个重要特点是，他们大多是人口基础。这意味着，他们是代表的个人设置免疫细胞和分子，它们完成既定的作用，组成，于我们而言，模式识别。如果我们重述3免疫程序审查，阴性选择，克隆选择，免疫网络，他们都以个人的人口数和依靠承认一个集P的模式。负选择算法来定义非我模式设置的探测器;克隆选择复制，交互，并选择自细胞能识别的非我集;和免疫网络维护一个个人设置，作为一个网络连接，以连接自我与非我。 首先考虑的二进制海明形状空间的情况下，这是最广泛的应用。有可以在之间的匹配或相邻元素和一个M的最简单的情况元素亲和力度测定雇几个表达式是简单地计算这两个因素之间的海明距离(1)。另一种方法是寻找一个的r -连续位序列，如果对r连续数之间的字符串匹配一个给定的阈值比大，然后承认说，已经发生。作为最后的办法是这里提到的，我们可以描述亨特亲和措施，由公式(2)给出。这最后一种方法的优点是它有利于比赛的互补序列，从而为字符串之间的属性(模式相似区域搜索)。 L1ifpm,,iiDwhere,,, (1) ,,,,H0otherwise,1i, liDD,，2 (2) ,Hi 在欧几里德形状空间的情况下，欧氏距离可以用来评估系统之间的任意两个组件的亲和力。如曼哈顿距离其他方法也可聘用。 请注意，所有的方法描述基本上依赖于确定字符串之间的匹配。但是，也有在文献中， 15 信息工程学院-软件工程 要考虑到，如每个抗体的模式匹配的数字，其他方面的客户资料安全。 4 免疫系统进行的一项调查显示,对模式识别方法 人工免疫系统的应用是广阔的,从机器学习机器人自主导航。本节将回顾一些工程应用的AIS文学的模式识别领域。这个理论是提供一个向导的文学和一个简短的描述的应用范围的算法。这个部分分为两个部分便于理解:1)计算机安全,2)其他方面的应用。保护的问题(或计算机网络计算机病毒,未经授权的用户)等等,构成了一个丰富的研究领域中,模式识别系统。由于是吸引人的直觉的隐喻的人工免疫系统的建设来检测计算机病毒,已经有了极大的兴趣,从计算机科学社群这个特定的应用。使用的消极和克隆选择算法已被广泛的测试。前者,因为它是一种内在的异常变化)检测系统的具体情况,制定了一个模式识别装置。后者,克隆选择算法,已被应用于连接到否定选择由于其学习能力。其他更多的经典模式识别的任务,如字符识别和数据分析,并研究了人工免疫系统内。 5 免疫系统和神经网络作为模式识别方法 类似于使用人工神经网络方法,执行模式识别与人工免疫系统通常包括三个阶段:1)定义的方式表达;2)改编(学习或进化)系统,确定了一套典型数据;3)应用系统识别一套新的模式(可能包含模式用于自适应阶段)。 摘要提出了三种免疫算法(阴性选择、克隆选择、免疫网络),再加上造型模式识别的过程中所描述的免疫系统,在第三节,这一节将人工免疫系统和神经网络相比聚焦模式识别中的应用。讨论将基于计算等方面的基本成分,适应机制等。常见的神经网络作为模式识别方法将被考虑,如单和多层感知器,联想记忆,自我组织网络。所有这些网络是集的人工神经元的单位;他们适应环境通过学习(或存储)算法,他们可以有他们的建筑动态适应连同权重,他们已经基本知识储存在连接优势。 成分:基本单元的免疫系统是一个属性字符串s(连同其连接的网络模型,在适当的环境条件下，这个字符串的可能与免疫细胞和分子。在人工神经网络的基本单位,是一种人工神经元所构成的一个总结,活化作用,连接的优点,活化阈值。而人工神经元通常处理元素,属性字符串的免疫细胞和分子的代表是信息存储和加工的零件。 位置的组成部分:在免疫网络模型,细胞与分子动力学行为通常试图模仿或阻止环境。这种方式,网络元素将根据环境刺激。与免疫网络模型,安妮有他们的神经元放置在网络上的位置固定预定义。一些神经网络模型也采用固定的邻居模式的神经元。如果一个网络模式的连接不采用人工免疫系统,每个元素将会有一个位置的变化动态,人口。同样,一个动态性过程可能允许介绍和/或消除特定的单位。 结构:在消极和克隆AIS、元件通常是围绕矩阵的个人技能或代表。这些矩阵可能有固定或变尺寸。在人工免疫网络和人工神经网络的组件的人口都是相互连通,围绕的模式。人工免疫网络架构,通常的空间分布,而抗原代表安通常有先验架构、和权重由环境确定。 记忆:属性字符串代表剧目的免疫细胞和分子,他们各自的编号,构成了大部分的知识包含在一个人工免疫系统。再者,参数,如亲和力阈值的一部分也可考虑的记忆,我在学习法语。在人工免疫网络模型的优点,也在单位联系信息,即内生和外生之间的相互作用,他们量化的AIS自己和环境。在大多数情况下,内存内容可检索的并分发出去。在标准(最早的)神经网络模型, 16 信息工程学院-软件工程 知识是储存在连接的优点,只有个别神经元。在更雄厚的策略,如建设性和修剪的算法和网络的自适应参数,最后用神经网络层数,连接和形状的各自的活化作用也部分的网络知识。记忆是通常自我检索或内容可检索的,并分发出去。 适应:适应通常指变更或调整,在结构和行为的一个制度,使其格局的回应其他的系统组成和对环境的变化。尽管演化和学习过程所涉及的改编,有一种概念上的区别。可以被看作是一个进化的基因组成的变化的人口在历代个体。这是由于自然选择会对个体之间的遗传变异。相反,学习可以被看作是一个持久的改变行为所导致的工作经验。虽然目前这两种类型的免疫系统可以适应、学习和演化、类主要通过学习适应过程。 可塑性和多样性主要是:两种过程:招募新组件系统和消除无用的要素,从系统作为结果,对动态性系统的结构可以更适当地适应环境的能力,其搜索(多样性增加。此外, 动态性减少冗余系统内消除无用的部件。在免疫算法的动态性相当于一个连续的插入和消除基本元素(细胞的分子组成的系统。在变化的动力学方面相当于修剪和/或插入新连接,单位,并在网络层。 与其他组件之间的相互作用中产生细胞和分子通过识别(AIS)的属性字符串匹配由细胞受体(其他属性字符串)。在免疫网络模型、电池通常加权联系,让他们与辨识,并承认在其他的细胞。这些权值可以是一些刺激性或压制表示与其他细胞相互作用的程度。人工神经网络是由一组(或设置)相互关联的神经元的连接优势承担任何正面或负面的价值观,显示一个兴奋或抑菌活性。在与其他神经元网络发生这些连接优势,显在单一神经元接受和处理输入的环境(或网络神经元)在同一或其他层。个别神经元也会收到一个输入从本身。 与环境的互动模式识别应用:环境是被描绘成一个套输入模式应该学、认可、和/或机密。在法语,一个属性字符串代表遗传信息的免疫细胞和分子。这根绳子比与模式,从环境。如果有一个明确的抗原人口是公认的(套模式),全部或部分抗原可以呈现给我们的全部或部分,我在学习法语。在最后阶段的学习或认可,每个组成部分的免疫系统会承认输入模式。人工神经元连接输入信号接收来自环境。这些信号进行处理,并将其与神经信息包含在人工神经网络,例如连接的强项。学习后,形成整个可能(大约)识别的输入模式。 形式:每个组成部分与其他细胞免疫分子相互作用的补充或在小周边地区,倡导一种参数亲和力的门槛。该阈值确定度之间的免疫细胞识别并输入模式。最新型号的神经元,包括一种偏见(或阈值)。该阈值的神经元活性测定,也就是说,它显示了如何将神经元活性敏感与输入信号。 稳健性:两者都非常强劲范式主要存在的数量或网络组件。这些元素,细胞、分子和神经元,能共同协作,和竞争力,来完成特定的任务。随着知识的分布在很多的系统组成、损坏或个人因素可能没有明显恶化的整体性能。双方都非常灵活,神经免疫和噪声的宽容。一个有趣的财产的免疫网络模型、反面选择算法,他们也自我识别,也就是说,他们学会认出自己。在免疫网络模型,细胞相互作用,通常这些互动优势的量化联系。在反面选择算法、自知之明是由储存资讯补充。 状态:在每步迭代计算,时间步或间隔,国家的AIS相浓度的免疫细胞和分子,和/或他们的亲缘关系。在免疫网络模型,连接各单位的优点也主要是当前状态的系统。在人工神经网络,激活神经元的输出级别确定系统的状态。注意这激活神经元的输出考虑连接数目的优点及各自的价值观、活化功能和网络的维度。 17 信息工程学院-软件工程 控制:任何免疫原理、理论或过程可以用来控制的类型的许多组件之间的交互的学习法语。作为例子,克隆选择可以被雇用来建造一个抗体能够识别一组曲目的抗原阴性选择模式,并可用于定义一套抗体(探测器)为识别异常的模式。鉴别和差分方程可用于控制一个人工免疫网络将与自身和环境。基本上,三种学习模式可以用来培养出一支:1)监督、 (2)非监督的强化学习。 普遍能力:免疫细胞和分子的案例,能够识别特定的模式,这不仅可以辨认,但也有特定的图案结构相关的模式。这个能力是由称为降糖药模仿的亲和力,可以利用阈值。任何模式躺在一个“邻居”的一个已知的模式可以被相同的组成部分,在已知的模式识别免疫。因此,一个组件的AIS一般可以认出其他元素的亲和力和比维生素e。除了降糖药,有些免疫学家推测,也可以杂交特定的抗体,在某种意义上,他们能够识别抗原的结构不同,相对而言,他们之间相互足够建立。基本上有两种方法可以达到满意的神经网络性能的数量减少了空间的尺寸参数,或减少了有效的维度。 在非线性问题: 非线性问题在AIS中出现在基本上使用激活函数定义程度的识别系统的两个部件之间的比例,对他们的亲和力。作为例子,一个息肉或一个简单的阈值函数也可以用于匹配。某些免疫网络模型用类似高斯曲线的功能的成熟和扩散的可能性很大程度上依赖程度的连通性的免疫细胞与当前的网络配置。非线性问题居住在人工神经网络的基本功能,每个神经元的活性。这一套操作数的非线性神经单元的结果在一个网络,具有极大的发展潜力进行非线性逼近和/或分类。 6 结束语 人工免疫系统的构成一个紧急生物动力计算范式。它是根据所提取的原则和隐喻的免疫系统以设计的计算工具解决复杂问题。事实上,主要的角色是免疫系统的识别哪些细胞、分子、组织属于机体并使他们与外国的元素。如果免疫系统并不是那么有效在这个自我/非歧视过程中,身体就没有问题的移植排斥反应的组织,例如。作为一种结果,这个伟大的能力来识别和排除特定模式(非)作为一个良好的灵感源泉,开发为机器学习和模式的计算模式识别方法。 在这一章三节课的人工免疫系统的执行模式识别算法,提出了一种消极的选择:1)否定选择;2)克隆选择,3)免疫网络模型,进行了综述。在否定选择、模式识别系统的设计是通过学习信息的补充设定模式是公认的一个全新的范例。克隆选择算法学会识别模式,通过一个进化程序。最后,免疫网络模型都是独一无二的,因为它们携带信息的模式是公认的,他们也有自己的知识,即:一种观念。所有的算法为基础的知识与人口分布组分之间的系统。 直觉和吸引人的隐喻的工程人工免疫系统的保护计算机和计算机网络病毒,未经授权的用户等,导致所谓的计算免疫学。大多数计算机免疫算法的具体案件,构成的人工免疫系统,是基于反面选择算法。在调查一章中,最具影响力的作品进行计算免疫学。此外,其他模型的应用,包括免疫网络和克隆选择算法的基础上,对其他类型的模式识别领域,如文字识别、数据分析、聚类和分类等问题进行了讨论。 本章的理论进行比较和人工免疫系统、神经网络模型对模式识别方法。等方面的基本单元构成每个系统,各自的类型的适应机制、不同类型的记忆方法,以及他们如何被强调,现在特 18 信息工程学院-软件工程 例能力。 还有几个在文学作品混合了神经网络与人工免疫系统;良好的评论中都能找到。虽然这并不包含在这里由于缺乏空间,这些作者坚信这两种方法都有很多利润。拉伸推测,这可能暗示小说模式,如很快就会出现人为的神经免疫系统。 本章的目的是为了服务的目的是介绍人工免疫系统的神经网络社区,还提供了基本的指导。提出的算法可以直接用于或通过改编来替代解决同一类型的模式识别问题作为神经网络,或补充他们的潜力。 19