Search

Gaussian process regression (GPR) has the ability to deal with non-linear regression readily, although the calculation cost increases with the sample size. In this paper, we propose a fast approximation method for GPR using both locality-sensitive hashing and product of experts models. To investigate the performance of our method, we apply it to regression problems, i.e., artificial data and actual hand motion data. Results indicate that our method can perform accurate calculation and fast approximation of GPR even if the dataset is non-uniformly distributed.

In recent years, most of mobile phones have a function of pedestrian navigation guidance. It was reported that users sometimes feel anxiety because of low accuracy of the position estimation especially in urban area and delay of information updating. In order to reduce the anxiety, a route planning algorithm is proposed in this study, which weighs user's difficulty (or easiness) of locating own current position as well as total physical distance of courses. The difficulty is estimated by valuation functions based on the "recognizability" and "visibility" of landmarks. An experimental study conducted in real situation using a prototype system to examine and refine the model for the optimal route planning. As the result, a modified model is proposed as a promising method of route planning for user's easy wayfinding.

This paper deals with a multi-agent path planning problem for tracking humans. Path of agents is planned based on the similarity between the prediction of the intensity of humans existing and the intensity of field of view of agents. Since the prediction is not always accurate, we proposed the an path planning method method where prediction length is varied based on the reliability of the prediction. We conducted computer simulation and results showed that our path planning method works well even under changing circumstances.

Recently, many humanoid robots have been developed and actively investigated all over the world in order to realize partner robots which coexist in the human environment. For such robots, high communication ability is essential in order to naturally interact with human. Among them, an android robot developed by Ishiguro et al.[1] has a quite similar appearance to a human. Therefore generating natural motions and behaviors is very important for the android since unnatural motion of the android may give worse impression than other robots as “uncanny valley” [4] insists. Noma et al. showed that 70% of human subjects were unable to distinguish the android from a human by observing it for 2 seconds when a small human-like fluctuations which corresponds to breathing and blinking were added to its static posture[5]. They also showed that 77% of subjects recognized the android as a robot when it was standing still. This indicates that small difference in behaviors of the android can give large difference in the impression on the android.

Controlling a highly dynamics and unknown system by existing control methods would be difficult because of its complexity. Recent biological studies reveal that animals utilize biological fluctuations to achieve adaptability to the environment and high flexibility. In this paper, we propose a simple, but flexible control method inspired by a biological adaptation mechanism. The proposed method is then applied to control robotic arms. The results of simulation indicated that our proposed method can be applied well to the control of a robot with multi-DOF.

The most important countermeasure against earthquake risk is to have all structures vulnerable enough for the possible earthquake load. In this regard, an early warning system should be installed to reduce the possibility of earthquake disaster. An early warning system is required mainly to issue an alarm to have a time margin for evacuating or shutting down key facilities, and not to determine exact earthquake parameters. Thus the early warning system must be realized independently and the government and other public authorities must release accurate earthquake information immediately.

There are two kinds of earthquake alarm. One is an “On-site Alarm” which is based on observations at the site of the objects under threat. The other is a “Front Alarm” which is the alarm based on observations near the epicentral area to spread the warning to potentially damaged areas. For each, there are two further kinds of alarm; one being the alarm exceeding a preset level (S-wave Alarm or Triggered Alarm), the other being the alarm during the preliminary motion (P-wave Alarm).

The scientific enterprise depends on the effective transfer of knowledge from creator to user. Recently the rate of scientific knowledge production is overwhelming the ability for researchers to process it. Semantic web technologies may help to handle this vast amount of scientific knowledge. However, automatic computerized techniques that extract semantics from natural language text for use in matching with the requests of knowledge seekers achieve only mediocre results. Clearly, semantic descriptions of expert knowledge that are constructed by the knowledge creators themselves will be more accurate. We report an approach and software implementation of a knowledge sharing platform based on semantic web technologies, called EKOSS for expert knowledge ontology-based semantic search, that helps knowledge creators construct semantic descriptions of their knowledge. The EKOSS system enables knowledge creators to construct computer-interpretable semantically rich statements describing their knowledge with minimal effort and without any knowledge of semantic web technologies.

In this article, we propose a feature extraction technique for decision-theoretic planning problems in partially observable stochastic domains and show a novel approach for solving them. To maximize an expected future reward, all the agent has to do is to estimate a Markov chain over a statistic variable related to rewards. In our approach, an auxiliary state variable whose stochastic process satisfies the Markov property, called internal state, is introduced to the model with the assumption that the rewards are dependent on the pair of an internal state and an action. The agent then estimates the dynamics of an internal state model based on the maximum likelihood inference made while acquiring its policy; the internal state model represents an essential feature necessary to decision-making. Computer simulation results show that our technique can find an appropriate feature for acquiring a good policy, and can achieve faster learning with fewer policy parameters than a conventional algorithm, in a reasonably sized partially observable problem.

There has been a problem called "exploration-exploitation problem" in the field of reinforcement learning. An agent must decide whether to explore a better action which may not necessarily exist, or to exploit many rewards by taking the current best action. In this article, we propose an off-policy reinforcement learning method based on a natural policy gradient learning, as a solution of the exploration-exploitation problem. In our method, the policy gradient is estimated based on a sequence of state-action pairs sampled by performing an arbitrary "behavior policy"; this allows us to deal with the exploration-exploitation problem by handling the generation process of behavior policies. By applying to an autonomous control problem of a three-dimensional cartpole, we show that our method can realize an optimal control efficiently in a partially observable domain.

Motivated by the perspective that animals’ rhythmic movements such as locomotion are controlled by neural circuits called central pattern generators (CPGs), motor control mechanisms by CPG have been studied. As an autonomous learning framework for a CPG controller, we previously proposed a reinforcement learning (RL) method called the CPG-actor-critic method. In this article, we propose a natural policy gradient learning algorithm for the CPG-actor-critic method, and applied our RL to an automatic control problem by a biped robot simulator. Computer simulations show that our RL makes the biped robot walk stably on various terrain.

An earthquake is a devastating natural disaster which can cause huge amount of casualties and extensive damage within a period of a few minutes. UrEDAS (Urgent Earthquake Detection and Alarm System) is an intelligent and unique alarm system. This system comprises of a single station including three component seismometers and does not need any extra organisation for networking. Immediately after the arrival of a P-wave, it provides information on magnitude, location and the depth of the earthquake. The damage potential is then estimated by relating the magnitude, epicentral distance and depth within three seconds after P-wave detection.

It has been widely recognized that systematic and continuous monitoring of earthquakes is necessary for early warning. On the other hand, investigation of the durability of surface ground and structures should always have priority, when estimating the damage caused by destructive earthquakes. To estimate the vulnerability of both surface ground and structures, Naka-mura (1997) proposed an effective approach called the K-value method. In this method, the deformation characteristics are approximated based on the amplification and the natural frequency information determined from microtre-mor measurements. These K-values, together with the acceleration are then used to calculate the strain of the ground and structures. Naka-mura (1997) verified this approach for several real earthquake cases by comparing estimated and actual damage. It was proven that it is possible to effectively estimate actual damage by combining continuously monitored earthquake motion with investigated characteristics of ground and structures. This paper addresses the formulation of the Kd value (K-value for the derailment of trains), proposing the new characteristic parameters PI (P-wave Index) and DI (Damage Intensity) proposing to help define the destructiveness of the earthquake and effectively realize the P-wave alarm system. During an earthquake, as soon as P-waves arrive, the DI function can be calculated in real time. This new characteristic DI-value (DIV) has a clear physical meaning and agrees very well with the instrumental seismic intensity of the Japan Meteorological Agency (Ijma)- The relation between DIV and IJMA can be approximately given as IJMA = D1v-0.6. This enables one to estimate the instrumental seismic intensity from DIV. Using this approach, it has been proved that, as with K-values, PI- and DIv-values can be used for early warning studies. In addition, actual damage can be accurately estimated and effective response can be decided on quickly.

This paper studies the reinforcement learning (RL) method for central pattern generators (CPG) that generates stable rhythmic movements such as biped locomotion. RL for biped locomotion is very difficult, since the biped robot is highly unstable and the system has continuous state and action spaces with a high degree of freedom. In order to deal with RL for CPG, we propose a new RL method called the CPG-actir-critic method. We applied this method to the RL for the biped robot. The computer simulation showed that our RL method was able to train the CPG such that the biped robot walk stably.

This paper examines an interval ordered structure in risky decision making and proves the existence of expected utility with a threshold function. Our interval ordered structure shows that the threshold function is a nonnegative linear functional. We also explore a special structure which gives a nonnegative constant threshold function.

This paper deals with a descriptive model to account for various paradoxes (e.g. Allais paradox) which violate the von Neumann-Morgenstern expected utility theory. Extending the prospect theory of Kahneman-Tversky we propose a “measurable value function under risk” which is a two-variable function of outcome and probability. The descriptive model proposed in this paper could properly account for Allais paradox (certainty effect), reference effect, and the phenomena of insurance and gambling. If we eliminated the risky situations from our model, we could obtain the conventional model of measurable value function under certainty as a special case.

Although EEG arousal induced by stimulation of brain stem activating system is now a familiar phenomenon to us, and close relationship between consciousness and functional states of this structure has been established, much is left unknown about the details of changes in the level of cortical excitability induced by arousal stimulation.