Catálogo de publicaciones - libros

An adaptive critic design is used in the nosiheptide fermentation process to solve the intractable optimization problem. The utility function is defined as the increment of biomass concentration at the adjacent intervals. The state variables are chosen as the biomass concentration, the substrate concentration, the dissolved oxygen concentration and the inhibition concentration. The decision variables are chosen as the temperature, the stirring speed, the airflow and the tank pressure. The adaptive critic method determines optimal control laws for a system by successively adapting the critic networks and the action network. The simulation shows at the same initial conditions this technique can make the fermentation shorten 6 hours.

This paper mainly discusses a generic scheme for on-line adaptive critic design for nonlinear system based on neural dynamic programming (NDP), more exactly, an improved action-depended dual heuristic dynamic programming (ADDHP) method. The principal merit of the proposed method is to avoid the model neural network which predicts the state of next time step, and only use current and previous states in the method, as makes the algorithm more suitable for real-time or on-line application for process control. In this paper, convergence proof of the method will also be given to guarantee the control to reach the optimal. At last, simulation result verifies the performance.

We extend a reinforcement learning algorithm, REINFORCE [13] which has previously been used to cluster data [10]. By using base Gaussian learners, we extend the method so that it can perform a variety of unsupervised learning tasks such as principal component analysis, exploratory projection pursuit and canonical correlation analysis.

According to Hebb’s , the brain has the capability of function localization. On the other hand, it is suggested that the brain has three different learning paradigms: supervised, unsupervised and reinforcement learning. Inspired by the above knowledge of brain, we present a hierarchical learning system consisting of three parts: supervised learning (SL) part, unsupervised learning (UL) part and reinforcement learning (RL) part. The SL part is a main part learning input-output mapping; the UL part realizes the function localization of learning system by controlling firing strength of neurons in SL part based on input patterns; the RL part optimizes system performance by adjusting parameters in UL part. Simulation results confirm the effectiveness of the proposed hierarchical learning system.

A novel machine learning paradigm, i.e. enclosing machine learning based on regular geometric shapes was proposed. It adopted regular minimum volume enclosing and bounding geometric shapes (sphere, ellipsoid, and box) or their unions and so on to obtain one class description model and thus imitate the human “Cognizing” process. A point detection and assignment algorithm based on the one class description model was presented to imitate the human “Recognizing” process. To illustrate the concept and algorithm, a minimum volume enclosing ellipsoid (MVEE) strategy for enclosing machine learning was investigated in detail. A regularized minimum volume enclosing ellipsoid problem and dual form were presented due to probable existence of zero eigenvalues in regular MVEE problem. To solve the high dimensional one class description problem, the MVEE in kernel defined feature space was presented. A corresponding dual form and kernelized Mahalanobis distance formula was presented. We investigated the performance of the enclosing learning machine via benchmark datasets and compared with support vector machines (SVM).

In this paper we derive a simple reinforcement learning rule based on a more general form of REINFORCE formulation. We test our new rule on both classification and reinforcement problems. The results have shown that although this simple learning rule has a high probability of being stuck in local optimum for the case of classification tasks, it is able to solve some global reinforcement problems (e.g. the cart-pole balancing problem) directly in the continuous space.

In electric power system, long term peak load forecasting plays an important role in terms of policy planning and budget allocation. The planning of power system expansion project starts with the forecasting of anticipated load requirement. Accurate forecasting method can be helpful in developing power supply strategy and development plan, especially for developing countries where the demand is increased with dynamic and high growth rate. This paper proposes a peak load forecasting model using relevance vector machine (RVM), which is based on a probabilistic Bayesian learning framework with an appropriate prior that results in a sparse representation. The most compelling feature of the RVM is, while capable of generalization performance comparable to an equivalent support vector machine (SVM), that it typically utilizes dramatically fewer kernel functions. The proposed method has been tested on a practical power system, and the result indicates the effectiveness of such forecasting model.

The forecasting models for stock market index using computational intelligence such as Artificial Neural networks(ANNs) and Genetic programming(GP), especially hybrid Immune Programming (IP) Algorithm and Gene Expression Programming(GEP) have achieved favorable results. However, these studies, have assumed a static environment. This study investigates the development of a new dynamic decision forecasting model. Application results prove the higher precision and generalization capacity of the predicting model obtained by the new method than static models.

All the factors that influence the rolling force are analyzed, and the neural network model which uses the back propagation (BP) learning algorithm for the calculation of rolling force is created. The initial network’s weights corresponding to the input material grades are taught by the traditional theoretical model, and saved in the database. In order to increase the prediction accuracy of rolling force, we use the measured rolling force data to teach the neural network after several coils of the same input material are rolled down.

This paper presents a hybrid modeling method which incorporates knowledge-based components elicited from human expertise into underlying data-driven neural-fuzzy network models. Two different methods in which both measured data and a priori knowledge are incorporated into the model building process are discussed. Based on the combination of fuzzy logic and neural networks, a simple and effective knowledge-based neural-fuzzy network model has been developed and applied to the impact toughness prediction of alloy steels. Simulation results show that the model performance can be improved by incorporating expert knowledge into existing neural-fuzzy models.