Use this URL to cite or link to this record in EThOS:
Title: Evolutionary and reinforcement fuzzy control
Author: Chowdhury, Mina Munir-ul Mahmood
ISNI:       0000 0001 3548 8489
Awarding Body: University of Glasgow
Current Institution: University of Glasgow
Date of Award: 1999
Availability of Full Text:
Access from EThOS:
Access from Institution:
Many modern and classical techniques exist for the design of control systems. However, many real world applications are inherently complex and the application of traditional design and control techniques is limited. In addition, no single design method exists which can be applied to all types of system. Due to this 'deficiency', recent years have seen an exponential increase in the use of methods loosely termed 'computational intelligent techniques' or 'soft- computing techniques'. Such techniques tend to solve problems using a population of individual elements or potential solutions or the flexibility of a network as opposed to using a rigid, single point of computing. Through use of computational redundancies, soft-computing allows unmatched tractability in practical problem solving. The intelligent paradigm most successfully applied to control engineering, is that of fuzzy logic in the form of fuzzy control. The motivation of using fuzzy control is twofold. First, it allows one to incorporate heuristics into the control strategy, such as the model operator actions. Second, it allows nonlinearities to be defined in an intuitive way using rules and interpolations. Although it is an attractive tool, there still exist many problems to be solved in fuzzy control. To date most applications have been limited to relatively simple problems of low dimensionality. This is primarily due to the fact that the design process is very much a trial and error one and is heavily dependent on the quality of expert knowledge provided by the operator. In addition, fuzzy control design is virtually ad hoc, lacking a systematic design procedure. Other problems include those associated with the curse of dimensionality and the inability to learn and improve from experience. While much work has been carried out to alleviate most of these difficulties, there exists a lack of drive and exploration in the last of these points. The objective of this thesis is to develop an automated, systematic procedure for optimally learning fuzzy logic controllers (FLCs), which provides for autonomous and simple implementations. In pursuit of this goal, a hybrid method is to combine the advantages artificial neural networks (ANNs), evolutionary algorithms (EA) and reinforcement learning (RL). This overcomes the deficiencies of conventional EAs that may omit representation of the region within a variable's operating range and that do not in practice achieve fine learning. This method also allows backpropagation when necessary or feasible. It is termed an Evolutionary NeuroFuzzy Learning Intelligent Control technique (ENFLICT) model. Unlike other hybrids, ENFLICT permits globally structural learning and local offline or online learning. The global EA and local neural learning processes should not be separated. Here, the EA learns and optimises the ENFLICT structure while ENFLICT learns the network parameters. The EA used here is an improved version of a technique known as the messy genetic algorithm (mGA), which utilises flexible cellular chromosomes for structural optimisation. The properties of the mGA as compared with other flexible length EAs, are that it enables the addressing of issues such as the curse of dimensionality and redundant genetic information. Enhancements to the algorithm are in the coding and decoding of the genetic information to represent a growing and shrinking network; the defining of the network properties such as neuron activation type and network connectivity; and that all of this information is represented in a single gene. Another step forward taken in this thesis on neurofuzzy learning is that of learning online. Online in this case refers to learning unsupervised and adapting to real time system parameter changes. It is much more attractive because the alternative (supervised offline learning) demands quality learning data which is often expensive to obtain, and unrepresentative of and inaccurate about the real environment. First, the learning algorithm is developed for the case of a given model of the system where the system dynamics are available or can be obtained through, for example, system identification. This naturally leads to the development of a method for learning by directly interacting with the environment. The motivation for this is that usually real world applications tend to be large and complex, and obtaining a mathematical model of the plant is not always possible. For this purpose the reinforcement learning paradigm is utilised, which is the primary learning method of biological systems, systems that can adapt to their environment and experiences, in this thesis, the reinforcement learning algorithm is based on the advantage learning method and has been extended to deal with continuous time systems and online implementations, and which does not use a lookup table. This means that large databases containing the system behaviour need not be constructed, and the procedure can work online where the information available is that of the immediate situation. For complex systems of higher order dimensions, and where identifying the system model is difficult, a hierarchical method has been developed and is based on a hybrid of all the other methods developed. In particular, the procedure makes use of a method developed to work directly with plant step response, thus avoiding the need for mathematical model fitting which may be time-consuming and inaccurate. All techniques developed and contributions in the thesis are illustrated by several case studies, and are validated through simulations.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Control systems