This paper describes a low-cost speech recognition system for controlling terminals and systems. The speech recognition system uses a compact speech recognition algorithm that is based on symbol matching for the five Japanese vowel sounds and a few consonant categories instead of conventional speech parameters. The algorithm has a significant advantage over conventional speaker-independent speech recognition algorithms that use complicated matching techniques; it reduces the memory required (by a factor from 4 to 200) for storing the reference templates and instruction sets and so can be implemented on low-cost processors. We implement this recognition algorithm on a fixed-point, 20-MIPS digital signal processor board with 10-k x 16-bit on-chip memory. Recognition experiments using 10 (20) Japanese city names indicate an accuracy of 96.2% (90.3%). By restricting the target vocabulary 20 words, the proposed speech recognition system can be integrated into many devices at Very low cost.
A hierarchical Bayesian approach is formulated for nonlinear time series prediction problems with neural nets. The proposed scheme consists of several steps : (i) Formulae for posterior distributions of parameters, hyper parameters as well as models via Bayes formula. (ii) Derivation of predictive distributions of future values taking into account model marginal likelihoods. (iii) Using several drastic approximations for computing predictive mean of time series incorporating model marginal likelihoods. The proposed scheme is tested against two examples; (A) Time series data generated by noisy chaotic dynamical system, and (B) Building air-conditioning load prediction problem. The proposed scheme outperforms the algorithm previously used by the authors.
In usual decision-making problems, attributes for an alternative are evaluated, and then, total evaluation for the alternative is done. Therefore, it is necessary to develop a suitable measure to integrate evaluations of given attributes. The difficulties in this process lie in : (i) dealing with the interaction among evaluations, (ii) taking into account some factors influencing an evaluation. However, any conventional literature on fuzzy measure has not discussed (ii) sufficiently. In this paper, P-measure which is defined on direct product space is proposed, to deal with (ii). Furthermore, we propose human measure, through t-conorm, which is possible to handle complexities (i) and (ii) simultaneously. Thus, the proposed human measure is found to be effective in decision-making problems.
In this paper, we discuss fast training of a support vector machine for pattern classification by extracting training data around the class boundaries, and then compare performance of the support vector machine with that of a fuzzy classifier with ellipsoidal regions. First, we discuss the architecture and the training method of the support vector machine. Then, we discuss how to extract boundary data from the training data. Next, we summarize the architecture of the fuzzy classifier and discuss a feature extraction method using a two-layer neural network. Finally, we compare performance of the support vector machine with the fuzzy classifier combined with the two-layer neural network for several benchmark data sets and demonstrate the effectiveness of the fast training method.
In the process industry, the closed-loop real-time optimization technology has become widespread with several commercial packages. However, the current technology mainly focuses on snapshot optimization, and it may not be used effectively in a time-variant process such as a reactor with catalyst degradation and a boiler with fouling growth. Ideally it is required to consider multi-period optimization that can handle future constraints and the total run-time cost on the process. The main difficulty in treating a multi-period process optimization problem resides in heavy computation due to the size of the problem. This paper proposes an efficient SQP algorithm that decomposes a QP subproblem into individual single-period QP problems by exploiting the particular structure of the multi-period optimization problem. Our numerical experiments indicate that the proposed SQP decomposition algorithm may be advantageous for problems with long time horizon.
In this paper we consider the scheduling problem of minimizing the maximum completion time (i.e., the makespan) for a two-machine robotic unit of flowshop type, in which each of n jobs is processed on the first machine and later on the second machine. There is an intermediate station between the two machines for intermediate operations such as washing, chip disposal, cooling, drying and/or quenching. If only permutation schedules are allowed (i.e., sequences of jobs on the two machines have to be the same), the problem can be solved in polynomial time, although non-permutation problem (i.e., the original problem) is NP-hard in the strong sense. It is already known that, if the optimal permutation schedule is used as an approximate solution for the non-permutation problem, it gives a maximum completion time within twice the optimal. In this paper, we present a different approximate algorithm which is based on a relaxation to the single-machine problem with delivery times, and show that its non-permutation schedule also gives a maximum completion time within twice the optimal. Moreover, we examine its approximation performance by means of numerical experiments, comparing with the optimal permutation scheduling.