Artificial Intelligence in Engineering最新文献

Design synthesis represents a highly complex task in the field of industrial design. The main difficulty in automating it is the definition of the design and performance spaces, in a way that a computer can generate optimum solutions. Following a different line from the machine learning, and knowledge-based methods that have been proposed, our approach considers design synthesis as an optimization problem. From this outlook, neural networks and genetic algorithms can be used to implement the fitness function and the search method needed to achieve optimum design. The proposed method has been tested in designing a telephone handset. Although the objective of this application is based on esthetic and ergonomic cues (subjective information), the algorithm successfully converges to good solutions.

Engineering design frequently relies on empirical data expressed in the form of non-dimensional correlations. These are almost always governed by applicability limits and the engineer is faced with the problem of choosing the right correlation that would provide design data with acceptable accuracy from a large number which are available. A knowledge based database system (KBDS) has been constructed which assists in the simple formulation of a jet impingement application based on which it retrieves and evaluates the relevant correlation from a database. Where the information in the database does not satisfy this specification the system uses knowledge of the application domain to either select suitable correlations for extrapolation or to modify the database query to select alternative information. The constraints which enable new correlations to be added or the knowledge in the network to be extended to include new geometries and flow conditions whilst maintaining the integrity are described. The operation of the KBDS has been demonstrated with a comprehensive database of correlations for the heat transfer due to the impingement of single and multiple air jets. This application provides typical engineering correlations and hence the techniques described are expected to be widely applicable.

Ensuring that the content of a rule-base, which is being encoded, is free from problems of consistency, completeness, and conciseness, is necessary to avoid any performance errors that might occur during consultation sessions with the rule-based system. In this paper we have described, formally, content verification of a specific type of rule-base using a digraph-based modelling approach. Through analytic formulations it is demonstrated that problems in the rule-base lead to the existence of certain properties in the digraph and various rule-base model representations that have been devised in this work. These properties, in turn, as is also shown through an example, can be examined for rule-base content verification.

Optimization methods are being applied to engineering problem solving with increasing frequency as computer hardware and software improves. The configuration of an optimization algorithm can make a significant difference to the efficiency of the solution process. This article examines the use of one such optimization strategy, branch and bound, for the solution of the classic knapsack problem. It is shown that the best configuration of the algorithm can be data dependent and hence that an ‘intelligent’ optimization system will need to automatically configure itself with the control knowledge appropriate to the problems the user is solving. A two-step approach is taken to configuring the algorithm. First, an analytical learning method, explanation based learning is used to derive a provably correct dominance condition for the knapsack problem. Second, the algorithm is configured with and without the condition, and subjected to a rigorous statistical test of performance, on the user's data, to decide which configuration is the best.

A hybrid fast training algorithm for feedforward networks is proposed. In this algorithm, the weights connecting the last hidden and output layers are firstly evaluated by the least-squares algorithm, whereas the weights between input and hidden layers are evaluated using the modified gradient descent algorithms. The effectiveness of the proposed algorithm is demonstrated by applying it to the sunspot and Mackey–Glass time-series prediction. The results showed that the proposed algorithm can greatly reduce the number of flops required to train the networks. The proposed algorithm is also applied to crowd estimation at underground stations and very promising results are obtained.

Multihop lightwave networks are a means of utilizing the large bandwidth of optical fibers. In these networks, each node has a fixed number of transmitters and receivers connected to a common optical medium. A multihop topology is implemented logically by assigning different wavelengths to pairs of transmitters and receivers. By using tunable lasers or receivers, it is possible to modify the topology dynamically when node failures occur or traffic loads change. The reconfigurability of logical multihop lightwave networks requires that optimal topologies and flow assignments be found. In this article, optimization of these logical topologies by genetic algorithms is investigated. The genetic algorithm takes topologies as individuals of its population, and tries to find optimal ones by mating, mutating and eliminating them. During the evolution of solutions, minimum hop routing with flow deviation is used to assign flows, and evaluate the fitness of topologies. The algorithm is tested with different sets of parameters and types of traffic matrices and the solutions are compared against histograms of random samples from the solution space. These tests show that the solutions found by the genetic algorithm are comparable with and in some cases better than those found by existing heuristic algorithms.

This paper reports on the application of Artificial Neural Networks (ANN) to long-term load forecasting. The ANN model is used to forecast the energy requirements of an electric utility. It is then compared to time series models. The comparison reveals that the ANN produces results that are close to the actual data. The ANN model is then used to forecast the annual peak demand of a Middle Eastern utility up to the year 2006. The results compare favorably with the utility’s forecast.

Autonomous vehicles must be able to navigate freely in a constrained and unknown environment while performing a desired task. To increase its autonomy, a vehicle must be provided by sophisticated software navigators. Traditionally, navigators build a convenient model of the vehicle's environment and plan feasible paths by reasoning about what actions must be performed to control the vehicle in that environment. This paper presents a genetic algorithm for adaptive navigation of a robot-like simulated vehicle. The proposed algorithm evolves feasible paths by performing an adaptive search on populations of candidate actions. The performance of the algorithm is demonstrated on problems with vehicles moving in two-dimensional grids and compared with that of a simple greedy algorithm and a random search technique.

Two of the well-known recurrent neural networks are the Elman network and the Jordan network. Recently, modifications have been made to these networks to facilitate their applications in dynamic systems identification. Both the original and the modified networks have trainable feedforward connections. However, in order that they can be trained essentially as feedforward networks by means of the simple backpropagation algorithm, their feedback connections have to be kept constant. For the training to converge, it is important to select correct values for the feedback connections, but finding these values manually can be a lengthy trial-and-error process. This paper describes the use of genetic algorithms (GAs) to train the Elman and Jordan networks for dynamic systems identification. The GA is an efficient, guided, random search procedure which can simultaneously obtain the optimal weights of both the feedforward and feedback connections.

This article describes a new adaptive fuzzy logic control scheme. The proposed scheme is based on the structure of the self-tuning regulator and employs neural network and genetic algorithm techniques. The system comprises two main parts: on-line process identification and fuzzy logic controller modification using the identified model. A recurrent neural network performs the identification and a genetic algorithm obtains the best process model and evolves the best controller design. The paper presents simulation results for linear and non-linear processes to show the effectiveness of the proposed scheme.