The authors describe a new learning algorithm for the multi-layer perceptron. The learning problem is stated formally as an optimization problem that is solved through a series of systematic approximations. The soluti...
详细信息
The authors describe a new learning algorithm for the multi-layer perceptron. The learning problem is stated formally as an optimization problem that is solved through a series of systematic approximations. The solution uses the moments of the training data to design the network. This procedure has several advantages, most importantly the reduction in training time. The algorithm is verified and compared to backpropagation. In a speech recognition experiment the total training time was reduced by more than 75% when compared to backpropagation.< >
The application of connectionist architectures for fast and robust online learning of dynamic relations used in robot control at the executive hierarchical level is discussed. The proposed connectionist robot controll...
详细信息
The application of connectionist architectures for fast and robust online learning of dynamic relations used in robot control at the executive hierarchical level is discussed. The proposed connectionist robot controllers use decomposition of robot dynamics. This method enables the training of neural networks on the simpler input/output relations with sigfnificant reduction of learning time. The other important features of these algorithms are fast and robust convergence properties because the problem of adjusting the weights of internal hidden units is considered as a problem of estimating parameters by the recursive least squares method and the extended Kalman filter approach. From simulation examples of robot trajectory tracking it is shows that when a sufficiently trained network is desired, the learning speed of the proposed algorithm is faster than that of the traditional backpropagation algorithms.< >
Training algorithms are introduced for single- and multiple-layered networks of Gaussian perceptrons. One characteristic of these algorithms is that they can guarantee that a network structure and the corresponding we...
详细信息
Training algorithms are introduced for single- and multiple-layered networks of Gaussian perceptrons. One characteristic of these algorithms is that they can guarantee that a network structure and the corresponding weights will be found for any arbitrarily given mapping relation of binary patterns. A number of computer simulation results are presented to demonstrate the performance of the proposed algorithms.< >
The authors present a method for the adaptive control of a robot arm based on a feedforward neural network. The method is based on the backpropagation algorithm. backpropagation is used within a learning by reinforcem...
详细信息
The authors present a method for the adaptive control of a robot arm based on a feedforward neural network. The method is based on the backpropagation algorithm. backpropagation is used within a learning by reinforcement framework instead of learning by teaching. A neural network is used to estimate the adaptive control law based only on an error signal resulting from the deviation between the desired position, velocity, and acceleration inputs to the robot inverse model and those generated by the robot system. The proposed method does not require any explicit parameter estimation of robot parameters. A cylindrical three-degree-of-freedom robot arm was simulated to demonstrate the control algorithm.< >
The authors compare the performances of a variety of algorithms in a reinforcement learning paradigm, including Ar-p, Ar-i, reinforcement-comparison (plus a new variation), and backpropagation of reinforcement gradien...
详细信息
The authors compare the performances of a variety of algorithms in a reinforcement learning paradigm, including Ar-p, Ar-i, reinforcement-comparison (plus a new variation), and backpropagation of reinforcement gradient through a forward model. The task domain is discrete multioutput functions. Performance is measured in terms of learnability, training time, and scaling. Ar-p outperforms all others and scales well relative to supervised backpropagation. An ergodic variant of reinforcement-comparison approaches Ar-p performance. For the tasks studied, total training time (including model and controller) for the forward model algorithm is 1 to 2 orders of magnitude more costly than for Ar-p, and the controller's success is sensitive to forward model accuracy. Distortions of the reinforcement gradient predicted by an inaccurate forward model cause the controller's failures.< >
The recurrent neural network is proposed for system identification of nonlinear dynamic systems. When the system identification is coupled with control problems, the real-time feature is very important, and a neuro-id...
详细信息
The recurrent neural network is proposed for system identification of nonlinear dynamic systems. When the system identification is coupled with control problems, the real-time feature is very important, and a neuro-identifier must be designed so that it will converge and the training time will not be too long. The neural network should also be simple and implemented easily. A novel neuro-identifier, the diagonal recurrent neural network (DRNN), that fulfils these requirements is proposed. A generalized algorithm, dynamic backpropagation, is developed to train the DRNN. The DRNN was used to identify nonlinear systems, and simulation showed promising results.< >
A fuzzy-based neural network (FBNN) model, which applies a one-pass algorithm is proposed. The theory of the FBNN model originates from embedding a fuzzy classification concept into a parallel neural network architect...
详细信息
A fuzzy-based neural network (FBNN) model, which applies a one-pass algorithm is proposed. The theory of the FBNN model originates from embedding a fuzzy classification concept into a parallel neural network architecture. Conventional neural networks, such as propagation using energy functions as learning principles, suffer from two major drawbacks, that of the local minimum problem and long training time. FBNN has the advantage of fast training, and avoids the local minimum problem. Experiments and comparisons between FBNN and some other neural network models are given. According to these results, FBNN shows stronger reliability on classification with respect to a probabilistic neural network, backpropagation, and a linear matching method.< >
The authors survey some of the fundamental aspects of neural networks that have been found crucial to their application to practical problems in diagnostics, modeling, and control. The analysis is a loosely connected ...
详细信息
The authors survey some of the fundamental aspects of neural networks that have been found crucial to their application to practical problems in diagnostics, modeling, and control. The analysis is a loosely connected collection of remarks on difficulties that have been encountered and the approach to dealing with them. Promising approaches now being explored and suggestions for future work are outlined. The issues raised concern the following: neural nets and engineering; training by higher order methods; sparse data and generalization; local representation networks; prestructured networks; scaling nodes; context switching; recurrent networks; neural controller development; and fusion of neural nets and fuzzy logic.< >
The author presents a connectionist natural language processing model, SAIL1, which uses a recurrent network topology to process English sentences. SAIL1 will build the sentence meaning representation incrementally, i...
详细信息
The author presents a connectionist natural language processing model, SAIL1, which uses a recurrent network topology to process English sentences. SAIL1 will build the sentence meaning representation incrementally, incorporating into the meaning only the information derived from words prior to the current word. The network is trained only on that part of the sentence meaning representing the data presented up to that time. Thus, the recurrent feedback is said to be realistic. SAIL1 has demonstrated its usefulness by successfully generalizing to create meaning representations for novel sentences.< >
暂无评论