National Technical University of Athens (NTUA)
Ανάπτυξη και εφαρμογές του αλγόριθμου S.Q.A. για την επίλυση προβλημάτων παραμετρικής βελτιστοποίησης της απόκρισης εύκαμπτων μη - γραμμικών δυναμικών συστημάτων
Abstract
dc:descriptionCompared to conventional, robust and bulky dynamic systems, lightweight flexible manipulators show remarkable advantages such as the possibility of performing greater variety of tasks in a remarkably wider scale, agility, faster response, manipulation of larger carried mass, the need for smaller drives reflecting to energy saving, easier transportation and installation and so on.The steering of modern lightweight manipulators usually sets problems that cannot be effectively resolved by the theoretical substratum within the framework of traditional analysis and design methodologies. These problems require the fulfillment of various and qualitatively diverse criteria and introduce decision variables in excess of those manageable by the classical methodologies. Further difficulties arise in the case of flexible arm systems, as it is usually necessary to adopt doubtful assumptions regarding the properties of the manipulated system, such as the invariance of the dynamic model parameters or the negligible influence of disturbance phenomena.In the afore-mentioned frame of requirements and conditions, it is very hard to derive distinct mathematical expressions for the degree of fulfillment of the particular criteria, as well as for the overall optimization problem formulation. Consequently, many researchers in the field resort to efficient non-conventional solution techniques, including the implementation of various parameter optimization methodologies. The current response optimization applications can be grouped in two main categories: the off-line design and the on-line applications. The first one, being to date the most popular, is the main research field of the present dissertation.The modern computing environments and the available software applications facilitate the development of increasingly powerful numerical tools for analyzing dynamic systems response. This way, the response calculation and through this, the evaluation of various test versions of a specific dynamic system and the consequent optimal selection can be realized easily, in the sense of not being time consuming and implying almost zero model manufacturing costs. The above can be performed for a practically unlimited range of excitation types and/or desired movements. To organize the relevant numerical procedure, one has to automatically generate various combinations of the design variables through an appropriate optimization algorithm.The response optimization of a dynamic system via the existing optimization methods necessitates parameterization of the physical problem (possibly by restating its original form), suitable formulation of a surrogate mathematical model (definition of objective function and equality and/or inequality constraints) and the choice of an appropriate optimization algorithm. The suitability of an algorithm for a category of problems depends on the form of the objective function, the possible existence of local optima, the number and the type (continuous or discrete) of the decision variables, the existence and the type of equality and inequality conditions, the computational cost for an analysis cycle, the required solution accuracy, etc.Flexible arms response optimization problems, typically, yield objective functions with singularities, hard to manipulate equality and inequality constraints and a multitude of localoptima combined with expensive analysis (each analysis cycle involves the solution of a non-linear system of differential equations in the time domain). Furthermore, for this type of problems, the analysis rarely provides information on the values of the derivatives of the objective function and the constraints with reasonable computational cost. The singular nature of the optimized function is further intensified in case that a series of pseudo-objective functions (e.g. penalty functions which incorporate the equality and/or inequality constraints in a weighted fashion) are utilized instead of the original one.The afore-mentioned peculiarities inherent to response optimization problems vitiate the classical deterministic optimization algorithms, which actually optimize a series of standard approximants of the (pseudo-)objective function and hence come up with unfortunate estimations for the optimum. This is typically the case when the mathematical model proposed does not succeed in effectively approximating the real objective function. Another significant complicating factor is the usual existence of many local optima, as the deterministic algorithms simply tend to converge to one of those closest to the optimization starting point.In the present dissertation a new optimization algorithm has been developed based on the specific nature of problems in parametric response optimization of non-linear dynamic systems including those of flexible manipulators. It is called “S.Q.A. algorithm” and relies on the deterministic model of Smoothed Quadratic Approximation surrogate function. This function is capable of exploiting unlimited amount of information coming from previous analyses, i.e. values of the objective function at points estimating the optimum during previous iterations. Another significant advantage of the S.Q.A function is its ability to overcome local singularities of the objective function. This way, the developed algorithm combines the main advantages of both deterministic optimization methodologies (swift and accurate optimum localization) and (semi-) stochastic ones (the ability to avoid sticking to local optima).The S.Q.A surrogate function is practically a compilation of smoothed quadratic approximations of the current (pseudo-) objective function. These approximations, which are explicitly presented in the third chapter, are formulated only upon zeroth order information and so that the existence and continuity of its first and second order derivatives is guaranteed even when the original objective function is discontinuous and/or its derivatives are discontinuous or do not exist. The formulation of the quadratic approximations mentioned above is always performed in relation with a specific starting point within the feasible domain and theoretically is capable of incorporating unlimited amount of available information (values of the objective function from previous, successful or not, estimations of the original global optimum).Every piece of available information (evaluation of the objective function for a specific combination of values of the decision variables) can be exploited (activated) at will, i.e. during any subsequent iteration and at any phase of the internal optimization procedure. The application of the proposed method in various cases showed that its effectiveness greatly depends on the quality of objective function approximations, which in turn depend on the multitude and distribution of the available information (analyzed points), the number of points and the rule by which they are selected to formulate any smoothed quadratic approximation, the formula for calculating the weights by which each point contributes to the approximationand a series of other adoptions which are detailed in chapters No. 2 and No. 3 of present thesis.The thesis at hand consists of eight (8) chapters. In the first chapter the most important response optimization methodologies are concisely presented in relevance with the available parameter optimization methodologies and the possibilities of intermediate mathematical problems formulation. Following, the primary, the secondary and the functional scopes of the dissertation are stated. The second chapter is about the step procedure to formulate the S.Q.A surrogate function and its basic properties, which are thereby proved (existence and continuity of partial derivatives up to the second order). We also set forth numerical data and graphical representations of the approximation for two characteristic cases of objective functions (one and two decision variables respectively). In the third chapter the theoretical background for second order optimization methods is briefly presented. Following, the structure of the basic version of the proposed algorithm is described. More specifically we analyze the main alternatives referring to the adaptability of S.Q.A algorithm to the specific nature of the optimization problem or a specific phase of the optimization procedure. The possible modifications of S.Q.A algorithm are explored in order to exploit potentially available information about the components of the gradient of the objective function. In forth section a parallel version of the S.Q.A algorithm is proposed which is fairly simple to implement and particularly useful for problems involving a relatively large number of decision variables and high computational cost per analysis. In chapter no. four many different test optimization problems are solved including functions with one, two or more local minima and with varying numbers of decision variables. In the fifth chapter the response of a half-car non-linear dynamic model is optimized when the vehicle run with invariable velocity over a constant inclination road irregularity, as well as over a sinusoidal one. The optimum response is defined according to the criterion of comfort (minimum absolute acceleration of vehicle center of gravity). A full set of diagrams is presented for both excitations cases corresponding to the most significant response variables (displacements, velocities, accelerations and adhesion forces).In chapter no. six a parameter optimization problem is formulated, including a series of equality and inequality constraints, equivalent to a classical optimal control problem of a discrete time system. Following, the detailed numerical results of parameter optimization procedure are presented.In the seventh chapter we select the parameter values of a non-linear control system applied to a single flexible arm for a wide range of values of the carried mass and for a complex sequence of angular displacement commands. Firstly, a field research of flexible arm systems is presented focusing on the most popular dynamic simulation methods, the utilized control systems and real – time response measuring techniques. Following, a full set of diagrams for the most significant response variables and the control signals is presented.In the eighth chapter the tip trajectory response of a non-linear dynamic system consisting of two jointed flexible arms is optimized for a wide range of values of the carried mass. In order to simulate the dynamic behavior of the system the equations of motion are greatly simplified by transfer to a Virtual Link Coordinate System (V.L.C.S). A full set of diagrams is also presented for the most significant response variables and the applied control torques.In chapter no. nine we evaluate the fulfillment of the primary, the secondary and the functional scopes of the dissertation, stated in the preclusive chapter. Finally, we express a series of propositions concerning the widening of S.Q.A algorithm applications as well as its potentials for further development.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2004
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Πετρίδης, Αθανάσιος
- Petridis, Athanasios
Subjects
dc:subject × 16- Βελτιστοποίηση απόκρισης
- Αλγόριθμοι Newton
- Βοηθητική συνάρτηση
- Εύκαμπτοι ρομποτικοί βραχίονες
- Βελτιστοποίηση μηδενικής τάξης
- Ομαλοποιημένη τετραγωνική προσέγγιση
- Response optimization
- Newton algorithms
- Surrogate function
- Flexible robot arms
- Zeroth order optimization
- Smoothed quadratic approximation
- Επιστήμες Μηχανικού και Τεχνολογία
- Επιστήμη Μηχανολόγου Μηχανικού
- Engineering and Technology
- Mechanical Engineering
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/16418