X.M. Gao

A modified Elman neural network model with application to dynamical systems identification

In this paper, an overview of the structure and learning algorithm of the Elman neural network is first presented. A modified Elman network is then proposed by adding new adjustable weights that connect the context nodes with output nodes. Convergence speed of the two network structures are compared. A parallel dynamic system identification scheme based on the modified Elman network is set up as well. Theoretical analysis and simulation results show that our improved neural network-based identification method has the advantage of identifying both linear and nonlinear dynamic systems without any prior knowledge of their orders and structures.

A hybrid optimization algorithm in power filter design

Clonal selection algorithm (CSA) is one of the most widely employed immune-based approaches for handling optimization tasks. Characterized with the similartaxis and dissimilation properties, Mind evolutionary computation (MEC) is a new evolutionary computation method. In this paper, we propose a hybrid optimization algorithm based on the principles of the CSA and MEC to search for the optimal parameters (values of inductor and capacitor) of a passive filter in the diode full-bridge rectifier. Simulation results demonstrate that our algorithm can acquire the optimal LC parameters within the given criteria for power filter design.

Genetic Algorithms-based Detector Generation in Negative Selection Algorithm

This paper proposes a genetic algorithms (GA) based detector optimization scheme in the negative selection algorithm (NSA). The NSA is a natural immune response inspired pattern discrimination method. In our scheme, the NSA detectors are optimized by the GA to occupy the maximal coverage of the nonself space so that they can achieve the best anomaly detection performance. Two numerical examples including the discriminant analysis of Fishers iris data are demonstrated to compare our new approach with a conventional detector generation method. Simulation results show that the former is more efficient than the latter for generating the NSA detectors

A New Harmony Search method in optimal wind generator design

The Harmony Search (HS) method is an emerging meta-heuristic optimization algorithm. However, like most of the evolutionary computation techniques, it does not store or utilize the useful knowledge gained during its search procedure in an efficient way. In this paper, we propose and study a new optimization approach, in which the HS method is merged together with the Cultural Algorithm (CA). Our modified HS method, namely HS-CA, has the feature of embedded problem-solving knowledge. The HS-CA is further employed in an optimal wind generator design problem, and it can yield a superior optimization performance over the original HS method.

Linguistic information feedforward-based dynamical fuzzy systems

In this paper, we first propose a linguistic information feedforward-based dynamical fuzzy system (LIFFDFS) in which the past fuzzy inference output represented by a membership function is fed forward locally with trainable parameters. Our LIFFDFS can overcome the common static mapping drawback of conventional fuzzy systems. We also give a detailed description of its underlying principle and general structure. Next, based on the gradient descent method, an adaptive learning algorithm for the feedforward parameters is derived. The proposed LIFFDFS is further employed in the prediction of time series. The well-known Box-Jenkins gas furnace data are used here as an evaluation example. Simulation results demonstrate that this new dynamical fuzzy system has the advantage of inherent dynamics and is, therefore, well suited for handling temporal problems, such as process modeling and control

A new CMAC neural network model with adaptive quantization input layer

We first discuss the structure, principle and learning algorithm of the cerebellar model arithmetic controller (CMAC) neural network model. A new adaptive quantization method based on competitive learning is then proposed to quantize the inputs of the CMAC according to the degree of variations of the approximated function. Theoretical analysis and simulation results show that with the input layer using this algorithm the CMAC can provide a more accurate and efficient approximation than the original model using equal-size quantization method.

Harmony Search Methods for Multi-modal and Constrained Optimization

The Harmony Search (HS) method is an emerging meta-heuristic optimization algorithm. In this chapter, we propose two modified HS methods to handle the multi-modal and constrained optimization problems. The first modified HS method employs a novel HS memory management approach to handle the multi-modal problems. The second modified HS method utilizes the Pareto-dominance technique, and it targets at the constrained problems. Several simulation examples are used to demonstrate and verify the effectiveness of our new HS methods.

Power prediction using an optimal neuro-fuzzy predictor

This paper presents a neuro-fuzzy predictor for received power level prediction in mobile communication systems. An important but difficult in designing such predictor is to the complexity of the predictor structure, i.e., the number of input nodes and the number of membership functions needed for each input node. We solve this problem by using the predictive minimum description length (PMDL) principle. This results in a predictor with excellent generalization capability. The optimized neuro-fuzzy predictor is then used for power prediction of simulated Rayleigh fading signals with 1.8 GHz carrier frequency. The results show that our optimized predictor can provide very accurate predictions of received signal power. Our neuro-fuzzy predictor is well suitable for applications where efficient compensation of fast fading and accurate power control are required.

Trajectory control based on a modified Elman neural network

We propose a modified Elman neural network (MENN)-based trajectory control scheme. Our approach consists of two stages: a MENN is first employed to identify the nonlinear plant to be controlled, and acts as the dynamical simulator of the plant after identification. The second MENN is then used to control the plant within the desired trajectory by backpropagating the control error through the simulator to update its weights. The application of this proposed method in the trajectory control of the inverted pendulum is illustrated. A simulation experiment demonstrates that our MENN-based trajectory control scheme can perform successful tracking without knowing prior knowledge of the plant.

Fast reinforcement learning algorithm for mobile power control in cellular communication systems

A fast reinforcement learning algorithm based on Mullers method is first proposed. This new algorithm converges much faster than the conventional approach, and therefore is more suitable to be used in on-line applications. The authors apply the fast reinforcement learning algorithm into the power control of cellular phones. The channel tracking error can be minimized in the mobile power control scheme. Simulation experiments demonstrate that the harmful deep fading is greatly compensated and the response overshoot is small.