[in Japanese]

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

In this paper, we propose a supervised self-organizing maps algorithm, named OrderSOM, which updates parameters using instance pairs with the relative order labels and then predicts the rank of unseen instances. Order SOM achieves order learning by adjusting the Best Matching Unit of the input instance based on its order labels. Through experiments conducted on synthetic dataset and real dataset, the effectiveness of OrderSOM was verified.

View full abstract

With the spread of mobile devices with positioning systems like GPS, user locations can be obtained in realtime with great accuracy. However, these realtime and accurate location information allows malicious people to identify users in the real space, which causes serious loss of anonymity. Though users can be anonymized with noises on their location, quality of location services would be also degraded. This paper proposes Virtual Scent, which is a novel method to obscure user locations without compromising k-anonymity as well as accuracy. Virtual Scent shows its users as ambiguous “scents” on a map, which can anonymize them. Though each scent has a noise on its location, they are positioned to be mixed up at a most likely location. We evaluate Virtual Scent with thorough simulation and experiments, which reveal its accuracy and practicality.

View full abstract

Towards building an intelligent robot, we have to create an autonomous mental development system that incrementally and speedily learns from humans, their environments, and the Internet. In this paper, we propose an ultrafast and online incremental attributes learning and transfer method using new SOINN; stand for Self-Organizing and Incremental Neural Network. We conducted a comparison experiment with previous methods. Based on these results, our proposed method can keep an equivalent recognition rate of an online method, and shorten the learning time and test time. And another advantage is to be able to use the Internet data, and be dened real valued attributes.

View full abstract

Fuzzy Q-learning has been studied that can treat a continuous state, since Q-learning treats only a discrete state. Dynamic Fuzzy Q-Learning (DFQL) has been proposed, where a new pair of state and actions is dynamically added to a given initial table of Q value. We propose a more flexible dynamic fuzzy Q-learning with facilities of tuning states of fuzzy sets and removing pairs of state and actions. We tune the center values and widths of fuzzy sets with TD (Temporal Difference) error of V value, which is evaluated value of states. We apply forgetting learning to fuzzy sets and V value and remove unnecessary fuzzy sets and unnecessary pairs of state and actions. We apply the method to the pursuit problem in a continuous environment.

View full abstract

A multi-robot system which is not controlled by any type of global controller and whose global behavior is emerged from the local interactions among robots or among robots and their environment is often called a swarm robotics system (SRS). While a number of researchers are interested in coordinating emergent behavior, only a few research projects deal with understanding and analyzing the collective behavior. To the best of our knowledge, no generally useful methods have been proposed in the case when robots have their physical bodies and as a result of that the congestion as well as collisions among robots occur in the simulated environment. In this paper, a novel method of analyzing the collective behavior incorporating the concept of behavioral sequence based on technique in ethology is proposed. In order to prove the potential power of the proposed method, a series of computer simulations of cooperative package-pushing problems is conducted to extract some features of collective behavior.

View full abstract