A neural network construction method for surrogate modeling of physics-based analysis.
详细信息
文摘
In this thesis existing methodologies related to the developmental methods of neural networks have been surveyed and their approaches to network sizing and structuring are carefully observed. This literature review covers the constructive methods, the pruning methods, and the evolutionary methods and questions about the basic assumption intrinsic to the conventional neural network learning paradigm, which is primarily devoted to optimization of connection weights or synaptic strengths) for the pre-determined connection structure of the network. The main research hypothesis governing this thesis is that, without breaking a prevailing dichotomy between weights and connectivity of the network during learning phase, the efficient design of a task-specific neural network is hard to achieve because, as long as connectivity and weights are searched by separate means, a structural optimization of the neural network requires either repetitive re-training procedures or computationally expensive topological meta-search cycles. The main contribution of this thesis is designing and testing a novel learning mechanism which efficiently learns not only weight parameters but also connection structure from a given training data set, and positioning this learning mechanism within the surrogate modeling practice. In this work, a simple and straightforward extension to the conventional error Back-Propagation BP) algorithm has been formulated to enable a simultaneous learning for both connectivity and weights of the Generalized Multilayer Perceptron GMLP) in supervised learning tasks. A particular objective is to achieve a task-specific network having reasonable generalization performance with a minimal training time. The dichotomy between architectural design and weight optimization is reconciled by a mechanism establishing a new connection for a neuron pair which has potentially higher error-gradient than one of the existing connections. Interpreting an instance of the absence of connection as a zero-weight connection, the potential contribution to training error reduction of any present or absent connection can readily be evaluated using the BP algorithm. Instead of being broken, the connections that contribute less remain frozen with constant weight values optimized to that point but they are excluded from further weight optimization until reselected. In this way, a selective weight optimization is executed only for the dynamically maintained pool of high gradient connections. By searching the rapidly changing weights and concentrating optimization resources on them, the learning process is accelerated without either a significant increase in computational cost or a need for re-training. This results in a more task-adapted network connection structure. Combined with another important criterion for the division of a neuron which adds a new computational unit to a network, a highly fitted network can be grown out of the minimal random structure. This particular learning strategy can belong to a more broad class of the variable connectivity learning scheme and the devised algorithm has been named Optimal Brain Growth OBG). The OBG algorithm has been tested on two canonical problems； a regression analysis using the Complicated Interaction Regression Function and a classification of the Two-Spiral Problem. A comparative study with conventional Multilayer Perceptrons MLPs) consisting of single- and double-hidden layers shows that OBG is less sensitive to random initial conditions and generalizes better with only a minimal increase in computational time. This partially proves that a variable connectivity learning scheme has great potential to enhance computational efficiency and reduce efforts to select proper network architecture. To investigate the applicability of the OBG to more practical surrogate modeling tasks, the geometry-to-pressure mapping of a particular class of airfoils in the transonic flow regime has been sought using both the conventional MLP networks with pre-defined architecture and the OBG-developed networks started from the same initial MLP networks. Considering wide variety in airfoil geometry and diversity of flow conditions distributed over a range of flow Mach numbers and angles of attack, the new method shows a great potential to capture fundamentally nonlinear flow phenomena especially related to the occurrence of shock waves on airfoil surfaces in transonic flow regime. Abstract shortened by UMI.).