Advances in Computational Intelligence

The following sections are included:

• Preface

• List of Contributors

• Contents

Granular computing is concerned with processing carried out at a level of coherent conceptual entities – information granules. Such granules are viewed as inherently conceptual entities formed at some level of abstraction whose processing is rooted in the language of logic (especially, many valued or fuzzy logic). The objective of this study is to discuss the role of Granular Computing and logic processing in the context of intelligent systems and demonstrate that both of them are organized into a single conceptual and computational framework. We provide a general overview of Granular Computing emphasizing a diversity of the currently available concepts and underlining their common features that make the entire pursuit highly coherent. The logic facet of processing is embodied in the framework of fuzzy logic and fuzzy sets that constitutes a consistent background necessary for representing and processing information granules. Several main categories of logic processing units (logic neurons) are discussed that support aggregative (and-like and or-like operators) and referential logic mechanisms (dominance, inclusion, and matching). We show how the logic neurons contribute to high functional transparency of granular processing, help capture prior domain knowledge and give rise to a diversity of the resulting models.

Linguistic dynamic systems (LDS) are dynamic processes involving mainly computing with words instead of numbers for modeling and analysis of complex systems and human–machine interfaces. The goal of studying LDS is to establish a methodology of design, modeling, and analysis of complex decisionmaking processes bridging the machine world in numbers and the human world in words. Specifically in this chapter, conventional dynamic systems are converted to different types of LDS for the purpose of verification and comparison. The evolving laws of a type-I LDS are constructed by applying the fuzzy extension principle to those of its conventional counterpart with linguistic states. The evolution of type-I LDS represents the dynamics of state uncertainty derived from the corresponding conventional dynamic process. In addition to linguistic states, the evolving laws of type-II LDS are modeled by a finite number of linguistic decision rules. Analysis of fixed points is conducted based on point-to-fuzzy-set mappings and linguistic controllers are designed for goals specified in words for type-II LDS. An efficient numerical procedure called α-cuts mapping is developed and applied to obtain extensive simulation results.

Data mining is an active area of research involving many large-scale machine learning issues. Ever-increasing dataset sizes and ever-larger problems have spurred research into efficient learning methods [4, 5, 7, 32, 37, 41, 45], including the traditional computer science method of divide and conquer. Distributed learning involves multiple learners working in a distributed fashion to solve the same problem. Distributed learning can lead to better scalability and is generally well-suited to the parallel and massively distributed architectures provided by the modern Internet. Slicing, or partitioning and learning, is a new learning algorithm that embodies the principles of distributed learning, not only to simplify a learning problem overall but also to simplify the individual learning tasks. It can be seen as a method in which the training set is partitioned into disjoint regions of feature space. These disjoint regions can be treated as a set of learning tasks that can be run in a distributed environment, without extensive communication. Classification takes place by assigning a test example to the classifier built from the corresponding region of feature space. Both the base learner and the partitioning strategy can be varied in this approach. Thus, slicing can be used as a general meta-learning technique for distributed learning. The slicing algorithm is examined with respect to a series of real-world datasets, including a biologically-motivated problem. Potential difficulties can occur with the slicing algorithm, however, solutions exist within the literature for mitigating the instability that arises when learning from less data. Slicing is demonstrated to be more accurate than using a single classifier, can reduce the individual learning task size and provides a mechanism to distribute the data mining task.

It is now commonly known that the margin-driven algorithms play a crucial role in statistical machine learning. In this chapter, we summarize our investigation on margin methods for both supervised and unsupervised learning problems. As for supervised learning, we first describe the nature of SVMs (support vector machines) in terms of margin-based generalization bound. Based on this nature, a complete framework of posterior probability support vector machines (PPSVMs) is proposed for weighted training samples using modified concepts of risk, linear separability, margin and optimal hyperplane. Within this framework, a new optimization problem for unbalanced classification problems is formulated and a new concept of support vectors is established. Furthermore, a soft PPSVM with an interpretable parameter ν is obtained which is similar to the ν-SVM developed by Schölkopf et al., and an empirical method for determining the posterior probability is proposed as a new approach to determine ν. The main advantage of a PPSVM classifier lies in that fact that it is closer to the Bayes optimal without knowing the distributions. To validate the proposed method, two synthetic classification examples are used to illustrate the logical correctness of PPSVMs and their relationship to regular SVMs and Bayesian methods. Compared with fuzzy support vector machines (FSVMs), the proposed PPSVM is a natural and analytical extension of regular SVMs based on the statistical learning theory. As for supervised learning, to extend the margin idea to unsupervised learning problems and establish a universal framework for one-class, clustering and PCA (principal component analysis) problems, an unsupervised learning problem with predefined threshold η is formally described and the intuitive margin is introduced. Then, one-class, clustering and PCA are formulated as three specific η-unsupervised learning problems. By defining a specific hypothesis space in η-one-class problems, the significant minimal sphere algorithm for regular oneclass problems is proved to be a maximum margin algorithm. Furthermore, some new one-class, clustering and PCA marginal algorithms can be achieved in different hypothesis spaces. Since the nature in SVM is employed successfully, the proposed algorithms have robustness, flexibility and high performance. To verify our formulation, some experiments are conducted. They demonstrate that the proposed framework is not only of theoretical interest, but they also have potentials in the family of practical unsupervised learning techniques.

In this chapter we present a generic, two-phase framework for solving constrained optimization problems using genetic algorithms. In the first phase of the algorithm, the objective function is completely disregarded and the constrained optimization problem is treated as a constraint satisfaction problem. The genetic search is directed toward minimizing the constraint violation of the solutions and eventually finding a feasible solution. A linear rank based approach is used to assign fitness values to the individuals in phase one. The solution with the least constraint violation is archived as the elite solution in the population. In the second phase the simultaneous optimization of the objective function and the satisfaction of the constraints are treated as a bi-objective optimization problem. We elaborate on how the constrained optimization problem requires a balance of exploration and exploitation under different problem scenarios and come to the conclusion that a non-dominated ranking between the individuals will help the algorithm explore further while the elitist scheme will facilitate in exploitation. We analyze the proposed algorithm under different problem scenarios using Test Case Generator–2 and demonstrate the proposed algorithm's capability to perform well independent of various problem characteristics. In addition, the proposed algorithm performs competitively with the state-of-the-art constraint optimization algorithms on eleven test cases which were widely studied benchmark functions in the literature.

Artificial neural network applications are numerous and rapidly growing over time for solving nonlinear complex problems. This chapter presents a hybrid training algorithm based on particle swarm optimization (PSO) and evolutionary algorithm (EA) for feedforward and recurrent neural networks. The application of the hybrid PSO-EA algorithm for training of a feedforward neural network as a board evaluator for the game Go and training of a recurrent neural network to predict the missing values from a time series are presented to show the potential of the algorithm. Results show the hybrid algorithm performs better than when the PSO or the EA algorithm is used individually as a training algorithm.

This chapter presents a novel approach to combining wavelet networks and multilayer feedforward neural network for fuzzy logic control systems. While most of the existing neuro-fuzzy systems focus on implementing the Takagi-Sugeno-Kang (TSK) fuzzy inference model, they fail to keep the knowledge structure that is critical in interpreting the learning process and providing insights to the working mechanism of the underlying systems. It is our intention to utilize individual subnets to implement decision-making process of the fuzzy logic control systems based on the Mamdani model. Center average defuzzification has seen its implementation by a neural network so that a succinct network structure is obtained. More importantly, wavelet networks have been adopted to provide better locality capturing capability and therefore better performance in terms of learning speed and training time. Offline orthogonal least squares method is used for training the wavelet subnets and the overall system is updated using the steepest descent algorithm.

Ant colony optimization (ACO) is a new meta-heuristic method based on the observation of real ant colonies' behavior. It provides a new way to solve many complicated optimization problems such as traveling salesman problem (TSP), quadratic assignment problem (QAP) and many others. A detailed overview on the principle of ACO and its applications are presented in this chapter. This survey should be helpful to the readers who are interested in ACO.

In this chapter, we study the problem of motif discoveries in unaligned DNA and protein sequences. The problem of motif identification in DNA and protein sequences has been studied for many years. Current popular algorithms for motif identification in protein sequences face two difficulties: high computational cost and the possibility of insertions and deletions. We propose a self-organizing neural network structure for solving the problem of motif identification in DNA and protein sequences. Our network contains several subnetworks with each performing classifications at different levels. The top level divides the input space into a small number of regions and the bottom level classifies all input patterns into motifs and non-motif patterns. Depending on the number of input patterns to be classified, several levels between the top level and the bottom level are needed to perform intermediate classifications. We maintain a low computational complexity through the use of the layered structure so that each pattern's classificationis performed with respect to a small subspace of the whole input space. Our selforganizing neural network will grow as needed (e.g., when more motif patterns are classified). It will give the same amount of attention to each input pattern and it will not omit any potential motif patterns. The definition of pairwise distance between patterns provided in this chapter can deal with up to two insertions/deletions allowed in a motif, while other existing algorithm can only deal with one insertion or deletion. Finally, simulation results show that our algorithm outperforms existing algorithms in certain aspects. In particular, simulation results show that our algorithm can identify motifs with more mutations than existing algorithms and our algorithm works well for long DNA sequences as well.

In this chapter, we discuss some computational problems that arise in the reverse engineering of protein and gene networks. We discuss the biological motivations, provide precise formulations of the combinatorial questions that follow from these motivations and finally describe the computational complexity issue, namely efficient approximation algorithms, for these problems.

Great improvement has been achieved in the on-line fingerprint sensing technology, automatic fingerprint recognition algorithms and fingerprint applications. Various fingerprint recognition techniques, including fingerprint acquisition, classification, enhancement and matching, have been developed. This chapter overviews recent advances in fingerprint recognition techniques and summarizes algorithms proposed for every step with special focuses on the enhancement of low-quality fingerprints and the matching of distorted fingerprint images. Both issues are believed to be significant and challenging tasks. We also discuss the common evaluation for fingerprint recognition algorithms: Fingerprint Verification Competition 2004 (FVC2004) and Fingerprint Vendor Technology Evaluation 2003 (FpVTE2003), based on which we can measure the performance of recognition algorithms objectively and uniformly. In addition, fingerprint recognition has been increasingly used to realize personal identification in civilian's daily life, such as fingerprint mobile phones, ID cards, fingerprint hard disks and so on. In this chapter, we introduce an application case: the fingerprint mobile phone which utilizes fingerprint recognition to ensure information security. It is composed of two parts: one is a front-end fingerprint capture subsystem, and the other is a back-end fingerprint recognition system based on smart phones.

The neuromuscular control system for posture and movement is a complex nonlinear system with many advanced features for adaptation and robustness in performance. Valuable lessons on how to design a versatile control system can be learned from investigation of functional structures of the system and strategies in adaptation to various tasks and environmental conditions. In this chapter we present two case studies on how the neuromuscular control system develops adaptive and predictive control strategies to achieve effective stability and performance improvement against perturbations. The first case is to maintain the upright posture against a strong yet predictable perturbation. The second case is to reach designated targets amid a strong pulling force perturbation. As results will demonstrate, a feedforward predictive control strategy is often developed to achieve efficient control with satisfactory performance, instead of energy demanding stiffness control.

Adaptive approximation methods estimate unmodeled nonlinear functions to improve tracking performance. This chapter investigates methods for the on-line estimation of functions that bound the achievable adaptive approximation accuracy. Whereas the results currently in the literature involved global forgetting, the adaptation laws for the estimated bounds herein involve local forgetting. The importance of local versus global forgetting is motivated in the text and illustrated with an example. Such localized bounds are useful for self-organizing adaptive approximators that adjust the structure of the function approximator during the operation of the system.

To solve the problems of conventional fuzzy inference, SIRMs (single input rule modules) dynamically connected fuzzy inference model is proposed in this chapter. For each input item, an SIRM is constructed and a dynamic importance degree is defined. The dynamic importance degree consists of a base value insuring the role of the input item throughout a control process, and a dynamic value changing with control situations to adjust the dynamic importance degree. Each dynamic value can easily be tuned based on the local information of current state. The model output is obtained by summarizing the products of the dynamic importance degree and the fuzzy inference result of each SIRM. The SIRMs process the input items dispersedly and the dynamic importance degrees express the control priority orders definitely. The present model is applied to several underactuated systems such as truck-trailer system, ball-beam system, and parallel-type double inverted pendulum system. The controller design approaches are given in detail. The simulation results indicate that the proposed model is effective even for very complex systems.

The following sections are included:

• Index