In this study, we discuss the principal carrier of speech information. Amplitude envelope of band-limited speeches was modeled by the proposed kernel function, producing three types of syntheses with modifications of the maximal part, the minimal part, and duration of the kernel function. Change of the envelope maxima only deactivated the temporal cortices in the brain by means of near-infrared spectroscopy (NIRS) measurements, which were in accord with the reduction of percentage correct in the dictation test. We hypothesized the integration of the envelope maxima played an important role in speech communication. As for NIRS methods, we regard the observation of hemoglobin maps as the most suitable index for the linguistic ability with following five advantages: (i) completely non-invasive measurements with a brief start-up, (ii) a bedside equipment with silent operation and continuous monitoring, (iii) hemodynamic observation in the cerebral association cortices, (iv) high flexibility in an experimental design because of tolerance for artifacts, and (v) low running costs without addition of supplies.

View full abstract

The present study is to address the extraction of speech information in the time course from filtered waveforms with a passband of one octave. Great many studies have found out significant information for the precise analysis of some resonances in the frequency domain, proposing many practical applications such as reproduction, recognition, compression for storage, and so on. Some neurophysiological studies, however, reported that frequency analysis in the auditory system was mechanically simple, generating a series of spikes, time-locked to the narrow-band vibrations on the basal membrane of the cochlea in the inner ear. Because the following extraction process is far from understanding, we synthesized speeches with increase of the minimal part of amplitude envelope. Conflict between subjective scores and quantization errors was observed when all the minimal parts in amplitude envelope were changed to 70% of the precedent envelope maxima in amplitude. Although quantization error is computationally 88.1%, which is larger than the allowance (>20-40%), the noise is subjectively inaudible. According to the forward masking effect, the results are interpreted as evidence that the envelope maxima and minima in the repeated structure correspond to the key information and the short-term connection with less perception, respectively. We estimate that inaudible structures amount to 70.4% for six channels belonging to a bandwidth of 80-5, 120Hz, and contribute to around 20% saving for storage.

View full abstract

In this article, we introduce a method of analysis of web readability based on Self-Organizing Maps (SOM) and pairwise comparison method. We focused on number of characters and line-feed-width among web contents for Web readability. Then we consider a threshold of readability for the specific age group using maps generated by SOM. We experimented to fifteen testers with method of pairwise comparison of Thurstone and Scheffe. As the result, we confirmed that in a certain range, appropriate numbers of characters are 32 characters and line-feed-width is 1.5 times. In addition, we confirmed that the readability had little influence of the difference fo browsing programs.

View full abstract

This paper relates the music recognition attempts using the symbolized particles store type neuron model. 12 pieces of musical sounds are prepared for experiments, and they are applied to the symbolized particles of the neuron model. They are also changed into codes by the model. The experimental result shows that the feature of musical sounds is expressed as to be in these codes. That is, it can be said that the application of the symbolized particle model to music recognition and to the feature extraction is very useful.

View full abstract

This paper presents a sentence-based information search method and its application to the assistance of summarization, considering and exploiting the expression features of the local assembly minutes to enable users to extract sentences of the specified type of information such as "policy description" and "opinion description." The assembly minutes are primarily the evidential record of the assembly activities, the preciseness is more weighed than the readability. However, the readability is an important issue to attract the interest of wider range of citizens, and our sentence-based search method contributes to that purpose. We show the application of the proposed method to the assistance of summarization. In contrast to the frequency-based keyword and sentence extraction as usually used in the automatic summarization methods, we show that more appropriate keywords are extracted from the members' question statements and the extraction of the sentences describing the administration's policies on those topics greatly improves the efficiency in making summaries that are important to attract the readers' interest.

View full abstract

Recently, Genetic Algorithm has been studied as an effective approach for large scale optimization problems. However, we have issues of early convergence and settings of many parameters in GA approach to actual applications. In order to deal with such issues, parameter free genetic algorithm(PfGA) and distributed genetic algorithm(DGA) are proposed. In this paper, we proposed distributed parameter free genetic algorithm(DPfGA) that keeps a parameter free characteristic and improve efficiency of the optimization. Besides the distributed construction of GA, we proposed the method varying the number of offspring adaptively in accordance with the current performance of the optimization. We showed effectiveness of the algorithm through application of the algorithm to TSP(Travelling Salesman Problem).

View full abstract

This research was done to clarify the meaning and the essential features of the HIC. At first, it is shown that as the order of the exponent in the equation of HIC is higher, the selected time interval becomes narrow, and the role of the exponent of 2.5 is a way to select the main peak from the noise. Next the data of the experiments using Articulated Total Body in JARI is shown, and the obtained actual accelerations are shown well fitted by Gauss function. Gauss fitting is well applied for almost all cases. Next we derived a simple formula to calculate the HIC based on Gauss function. As a result, HIC are expressed by only two parameters, i.e., amplitude A and distribution (sigma). The results will be important for the future works of the analysis based on the HIC. We show that the HIC strongly depends on the sharpness of the peak. The sharpness is determined by floor condition and protectors for head and body. It is concluded that helmet is very useful since it makes a long duration in collision, then it makes HIC lower than that without a helmet.

View full abstract

In order to realize an intelligent agent such as autonomous mobile robots, Reinforcement Learning is one of the necessary techniques in the behavior control system. However, applying the reinforcement learning to actual sized problem, the "curse of dimensionality" problem in partition of sensory states should be avoided maintaining computational efficiency. In multi-agent reinforcement learning such as the "pursuit problem". the problem is emerged owing to the number of collaborating allied agents and a prey agent, as well as the size of environments. We proposed a hierarchical modular reinforcement learning in order to deal with the dimensional problem and task decomposition. In this study, we focused on investigation of the learning performance of agent that represents the input states in relative coordinate system, considering application as mobile robots. We showed effectiveness of the proposed learning algorithm based on relative expressions with limited view through numerical experiments of the pursuit problem.

View full abstract

Reinforcement learning will be one of necessary techniques in the control system of the intelligent agent such as autonomous mobile robots. Fuzzy Q-learning is one of the promising approaches for implementation of reinforcement learning functions owing to its high ability of model representation. However, in applying fuzzy Q-learning to actual applications, the number of iterations for the learning also becomes huge as well as almost all Q-learning applications. Furthermore convergence performance is often deteriorated owing to its complicated model structure. In this study, implementation methods of fuzzy Q-learning were discussed in order to improve the learning performance of fuzzy Q-learning. The modular fuzzy model construction method based on fuzzy Q-learning was proposed in this paper. Multi-grain configurations of the modular fuzzy model are compared with parallel structured learning schemes. Through numerical experiments of the mountain car task and the Acrobot task, we found that the proposed construction of modular fuzzy model could improve the performance of fuzzy Q-learning.

View full abstract