There is a challenge ahead in the rail industry to accommodate increased demand. Time spent at the platform train interface (PTI) as passengers board and alight, rather than on the move, represents a limitation on system capacity. To overcome this, we propose RateSetter: an evolutionary optimiser that for the first time provides more effective PTI design choices based on passenger flow time and safety. An agent based passenger simulation model validated with CCTV footage is employed for fitness evaluation. The initial results provide guidelines not only for future PTI designs but also for retrofit designs to existing infrastructure, evaluating the effectiveness and diminishing returns of PTI features for the considered scenarios. Furthermore, it is observed that the proposed optimal PTI designs could significantly reduce the flow time for the cases examined. Results show that retro-fit designs could reduce the flow time in the range of 10%-35%.

The Common Spatial Patterns (CSP) algorithm is useful for calculating spatial filters for detecting event-related desynchronization (ERD) for use in ERD-based brain-computer interfaces (BCIs). However, basic CSP is a supervised algorithm suited only to two-class discrimination; it is unable to solve multiclass discrimination problems. This paper proposes a new method named the binary common spatial patterns (BCSP) algorithm to extend the basic CSP method to multi-class recognition. Our method arranges the spatial filters and Fisher classifiers in the form of a binary tree whereby N − 1 spatial filters and N − 1 Fisher classifiers are calculated for N class recognition. This is fewer than must be calculated in other methods (e.g. one-versus-rest, OVR). This makes the overall classification procedure less redundant. Simulation results show that BCSP has better performance than the OVR scheme and outperforms the three best teams in the 2008 BCI-competition.

In this paper, the effectiveness of three different operating strategies applied to the Fuzzy ARTMAP (FAM) neural network in pattern classification tasks is analyzed and compared. Three types of FAM, namely average FAM, voting FAM, and ordered FAM, are formed for experimentation. In average FAM, a pool of the FAM networks is trained using random sequences of input patterns, and the performance metrics from multiple networks are averaged. In voting FAM, predictions from a number of FAM networks are combined using the majority-voting scheme to reach a final output. In ordered FAM, a pre-processing procedure known as the ordering algorithm is employed to identify a fixed sequence of input patterns for training the FAM network. Three medical data sets are employed to evaluate the performances of these three types of FAM. The results are analyzed and compared with those from other learning systems. Bootstrapping has also been used to analyze and quantify the results statistically.

In this paper, an empirical study of the development and application of a committee of neural networks on online pattern classification tasks is presented. A multiple classifier framework is designed by adopting an Adaptive Resonance Theory-based (ART) autonomously learning neural network as the building block. A number of algorithms for combining outputs from multiple neural classifiers are considered, and two benchmark data sets have been used to evaluate the applicability of the proposed system. Different learning strategies coupling offline and online learning approaches, as well as different input pattern representation schemes, including the "ensemble" and "modular" methods, have been examined experimentally. Benefits and shortcomings of each approach are systematically analyzed and discussed. The results are comparable, and in some cases superior, with those from other classification algorithms. The experiments demonstrate the potentials of the proposed multiple neural network systems in offering an alternative to handle online pattern classification tasks in possibly nonstationary environments.

For transmission of digital data over a linear channel with additive white noise, it can be shown that the optimal symbol-decision equalizer is nonlinear. The Kernel Adaline algorithm, a nonlinear generalization of Widrow's and Hoff's (1960) Adaline, is capable of learning arbitrary nonlinear decision boundaries while retaining the desirable convergence properties of the linear Adaline. This work investigates the use of the Kernel Adaline as an equalizer for such transmission channels. We show that the performance of the Kernel Adaline approaches that of the optimal symbol-decision equalizer given by Bayes theory and further, still produces useful results when the additive noise is nonwhite. A description and preliminary results of an adaptive version of the Kernel Adaline are also presented.

In this paper, a study of the effectiveness of a multiple classifier system (MCS) in a medical diagnostic task is described. A hybrid network, based on the integration of a fuzzy ARTMAP and the probabilistic neural network, is employed as the basis of the MCS. Outputs from multiple networks are combined using some decision combination method to reach a final prediction. By using a real medical database, a set of experiments has been conducted to evaluate the performance of the MSC with different network configurations. The experimental results reveal the potential of the MCS as a useful decision support tool in the medical field.

In this work the kernel Adaline algorithm is presented. The new algorithm is a generalisation of Widrow's and Hoff's linear Adaline, and allows to approximate non-linear functional relationships. Similar to the linear adaline, the proposed neural network algorithm minimises the least-mean-squared (LMS) cost function. It can be guaranteed that the kernel Adaline's cost function is always convex, therefore the method does not suffer from local optima as known in conventional neural networks. The algorithm uses potential function operators due to Aizerman and colleagues to map the training points in a first stage into a very high dimensional non-linear “feature” space. In the second stage the LMS solution in this space is determined by the algorithm. Weight decay regularisation allows to avoid overfitting effects, and can be performed efficiently. The kernel Adaline algorithm works in a sequential fashion, is conceptually simple, and numerically robust. The method shows a high performace in tasks like one dimensional curve fitting, system identification, and speech processing.