Catálogo de publicaciones - libros

In this work, a general framework for developing learning rules with an added term (perturbation term) is presented. Many learning rules commonly cited in the specialized literature can be derived from this general framework. This framework allows us to introduce some knowledge about vector quantization (as an optimization problem) in the distortion function in order to derive a new learning rule that uses that information to avoid certain local minima of the distortion function, leading to better performance than classical models. Computational experiments in image compression show that our proposed rule, derived from this general framework, can achieve better results than simple competitive learning and other models, with codebooks of less distortion.

Reinforcement Learning (RL) is analyzed here as a tool for control system optimization. State and action spaces are assumed to be continuous. Time is assumed to be discrete, yet the discretization may be arbitrarily fine. It is shown here that stationary policies, applied by most RL methods, are improper in control applications, since for fine time discretization they can not assure bounded variance of policy gradient estimators. As a remedy to that difficulty, we propose the use of piecewise non-Markov policies. Policies of this type can be optimized by means of most RL algorithms, namely those based on likelihood ratio.

JABAT is a middleware supporting the construction of the dedicated A-Team architecture that can be used for solving variety of computationally hard optimization problems. The paper includes a general overview of the JABAT followed by a description and evaluation of the architecture designed by the authors with a view to solving RCPSP and MRCPSP instances. To construct the proposed system a number of agents, each representing a different optimization algorithm including local search, tabu search, as well as several specialized heuristics have been used. The system has been evaluated experimentally through solving a set of benchmark instances of the RCPSP and MRCPSP.

We developed a computational model of the mushroom body (MB), a prominent region of multimodal integration in the insect brain, and tested the model’s performance for non-elemental associative learning in visual pattern avoidance tasks. We employ a realistic spiking neuron model and spike time dependent plasticity, and learning performance is investigated in closed-loop conditions. We show that the distinctive neuroarchitecture (divergence onto MB neurons and convergence from MB neurons, with an otherwise non-specific connectivity) is sufficient for solving non-elemental learning tasks and thus modulating underlying reflexes in context-dependent, heterarchical manner.

Following the rapid development of Grid computing, Grid technology has been introduced into the manufacturing realm and is contemporarily being considered for the sharing of manufacturing resources. However current research in the subject-area is still immature and mainly focuses on conceptual framework development. Here a concrete performance-based Bayesian method for resource collaboration optimization in Extended Enterprise is proposed which improves and promotes research in applying Grid-thinking in inter-organizational manufacturing value chains. Based on the research background, problem statement, and the consideration of Bayesian learning, the method for probability dependency relationship modeling between the performance values of different manufacturing resource nodes in the Extended Enterprise is analysed; and is subsequently complimented by the development of an extended method for more general use. Finally, a system dynamics simulation model for the proposed method is established and the validity and effectivity of the suggested method is tested via a simple case study.

This paper presents a hybrid simulated-annealing (SA) algorithm for two-dimensional strip packing problem (2D SPP) where a set of small rectangular items has to be allocated on a larger stock rectangle in order to find a minimum height. A new recursive placement procedure is proposed and the procedure is combined with the SA algorithm. The hybrid-SA algorithm was tested on a set of benchmark problems taken from the literature. The computational results have validated the quality of the solutions and usefulness of the proposed hybrid-SA algorithm.

An algorithmic approach for handling linguistic values defined in the same linguistic variable is proposed. It can explicitly capture the differences of individuals’ subjectivity regarding linguistic values. The proposed approach can be employed as a useful tool for discovering hidden relationship among linguistic values. Thus, it provides a basis for improving the precision of knowledge acquisition in handling linguistic values. We apply the proposed approach to a collective linguistic assessment among multiple experts.

To improve the safety of drivers and walkers in a city, several traffic monitors are usually set on lanes. These traffic monitors can also improve the security of communities. In this paper, we integrate the so-called linear/circular consecutive--out-of-:F systems into our proposed traffic-monitor system. The objective is to find the optimal design of monitors under limited budget for the system. The main purposes of this paper are : (1) to propose a new traffic-monitor system, (2) to present an immune algorithm (IA) for the optimal design of traffic monitors, and (3) to report numerical results of various parameters by the proposed algorithm. It is shown that the proposed immune algorithm performs well for all test problems.

In wireless sensor networks where sensors are geographically deployed in 3D spaces, a mobile robot is required to travel to each sensor in order to download the data. The effective communication ranges of sensors are represented by spheres with varying diameters. The task of finding the shortest travelling path in this scenario can be regarded as an instance of a class of problems called Travelling Salesman Problem with Neighbourhoods (TSPN), which is known to be NP-hard. In this paper, we propose a novel approach to this problem using Estimation of Distribution Algorithms (EDAs), which can produce significantly improved results compared to an approximation algorithm.

Air target identification is an important issue in threat warning, airline security and surveillance. To obtain accuracy and reliability, the multisensor is used to give multiple sources information. Thus, an algorithm to fuse the information from the multisensor is needed. The (Dempster-Shafer) evidence theory is a generalization of Bayesian statistics. Evidential reasoning is suited to a range of decision-making activities. But it is invalid when dealing with conflicting probabilities. In this paper, a new weighted D-S combination rule is proposed to solve the conflicting management in the air target identification system. In the weighted method presented here, it is to modify evidences rather than to modify the combination rule. The rationality and effectiveness of the weighted method are evaluated by the target identification system.