This paper describes a human-agent interaction in which auser and a life-like agent mutually acquire the other's mind mappingthrough a mutual mind reading game. In these several years, a lot ofstudies have been done on a lifelike agent such as a Micro Soft agent, an interface agent. Through development of various life-like agents, amind like emotion, processing load has been recognized to play animportant role in making them believable to a user. For establishingeffective and natural communication between an agent and a user, theyneed to read the other's mind from expressions and we call the mappingfrom expressions to mind states mind mapping. If an agent and auser don't obtain these mind mappings, they can not utilize behaviorswhich significantly depend on the other's mind. We formalize such mutualmind reading and propose a framework in which a user and a life-likeagent mutually acquire each other's mind mapping. In our framework, auser plays a mutual mind reading game with an agent and they graduallylearn to read the other's mind through the game. Eventually we fullyimplement our framework and make experiments to investigate itseffectiveness.

View full abstract

Most previous studies about a mechanism of human language/meaning acquisition and its computational model did not take into account the instructor's adaptation to the learner. In this paper, we propose a mutual adaptive speech meaning acquisition model which can induce and utilize the instructor's adaptation, based on the observational results of meaning acquisition process through the actual interactions between human instructors and human learners. Specifically, this model is designed to achieve the following two competences: one is the competence to achieve a gradual learning which starts rather simple and easy learning task and then shift to more complex one; the other is to induce the instructor's adaptive behaviors by means of showing its learning level as the velocity of paddle motion, e.g., when the learning is going well, the paddle shows the fast action; otherwise, the paddle does the slow one. As a result of testing experiment to observe the interaction between participants and this model, although this model can not learn the meanings of participants' instructions because of the low accuracy of implemented sound recognizing system, this model can induce the participants' adaptive behaviors that the participants tried fitting their teaching strategies for this learning model.

View full abstract

In this paper, we propose a new method, called IOA, for variousbehaviors of communication robots. If robots behave according to ruledesigned a priori, it is too easy for human to predict the robot's behaviors. On the other hand, randomly selected actions lead chaosbehaviors, which prevents communication. IOA generates its own criteriafor robot's behaviors spontaneously through observation of otherrobots' behaviors and environment. In particular, the generatedcriterion is stable in short term but various in long term. Therefore, robots with IOA can generate consistent behaviors in short term, butcan change their rules over time. We implemented IOA on a humanoidrobot, then we confirmed it generates various criteria for behaviors.

View full abstract

In order to transmit facial expressions of the user using the face of a lifelike agent in agent-mediated distance communication, we have to realize facial expression mapping, which gives synthetic facial expressions of the agent appropriate for representing facial expressions of the user in communication. However, the criterion to determine which expression on the agent's face is the "same" as that on the user's face heavily depends on the subjective view of the user. This article discusses how to learn the facial expression mapping reflecting the user's subjective view from the interaction with the user for agent-mediated communication. In the discussion, we focus on the two problems: how to make the facial expression mapping realized by the system reflect the user's subjective view gradually through the interaction, and how to realize the facial expression mapping with the error less than the range acceptable for the user. We propose an approach that copes with these two problems by (1) acquiring positive examples and negative ones for the facial expression mapping reflecting the user's subjective view from the interaction with the user, (2) estimating the acceptable range of the error to realize the facial expression mapping, and (3) constituting RBF(Radial Basis Function) network that realizes the facial expression mapping within the estimated range of the error.

View full abstract

In this paper, we propose Multi User Learning Agent (MULA) to interact with various types of people effectively towards realization of Social Interaction. MULA is equipped with two functions for social learning. One is direct learning function by individualization of user parameters, and another is indirect learning function by past experience and the similarity between each users. MULA adopts an extended classifier system, which uses a classifier strength and a user profile for each people. Each person's classifier strength is enhanced by the interaction with the person and the past experience by similar type of people. We verified the effect of MULA by experimenting in interaction with six human beings using AIBO ERS-220A. It confirmed that this system promoted learning efficiently to two or more persons.

View full abstract

[in Japanese]

View full abstract

[in Japanese]

View full abstract

In this paper, modellng technique is proposed. This modeling technique is based on concept of "soft computing". "soft computing" is an information processing technique for aiming at tractability, robustness, and low solution cost, by exploiting the tolerance for imprecision and avoiding the pursuit of excess precision. It is different from a conventional technique that uses elaborate algorithm and complex mathematical principle model. From this viewpoint, we also advocated an original modeling technique with reference for the process of vision cognitior, and have tried some application. The flow of this technique is as follows: at flrst transposes the characteristic of an object system to two or more ambiguous 2-dimensional geometric pattem information, transforms those pattems by leaming, finally combines them by the fuzzy technique, and performs the modeling of an object system. This technique was created by reference from the process of vision cognition: Detection of more global feature from local feature, and the parallelism of a lot of information expression and processing. We are calling this technique Pattern Infiormation Based Active Leaming Method (PBALM). PBALM have no mathematical formula and complicated algorithm, but just based only on simple operation. Moreover, as mentioned above, since it resolves object system into simple pattem information and treats as long as it is expressible by the patterrn, it can respond to large nonlinearity (robustness), Ieaming can also be performed with the intelligible form of modification of the pattem (tractability), and generally, there is little number of times of learning (low solution cost). In this paper, the idea of PBALM and the characteristic are described first. And the model of the air-conditioning jet wind velocity characteristic which is a nonlinear system is built. And it is shown that highly precise modeling is performed. Furthermore, the control problem of an inverted pendulum is taken up and it is shown that an effective control rule can be gained by PBALM using human's control data.

View full abstract

A Fuzzy Constraint Satisfaction Problem (FCSP) is an extension of the classical CSP, a powerful tool for modeling various problems based on constraints among variables. In traditional CSPs and FCSPs, values for the variables are chosen from discrete domains. However, this is often inconvenient when one wants to express real world problems. In this paper, we propose a new model that allows the variables to have a mixture of discrete and continuous domains. We show that this model, called HDFCSP (Hybrid Domain FCSP), can be solved by a new algorithm Spread Repair, an extension of the well-known iterative improvement algorithms. Experimental results on some test problems show that the algorithm has an ability of finding practical approximate solutions with high probability in a computation time much shorter than the traditional algorithms for the discrete-domain FCSPs.

View full abstract

Color transfer is an important problem in the field of non-photorealistic rendering research. We have developed and propose an approach that adopts interactive evolutionary computation to solve the problem. Conventional studies try to transform the colors and select only one output, which may have no relationship to the user's requirements. Our approach aims to obtain and optimal output for each user. We introduce dividing gene-coding and an adaptive probability for color selection to reduce user costs and obtain optimal solutions quickly. Our approach updates the probability for color selection, based on user's evaluation history. We will present and confirm the validity of our approach and demonstrate its' efficiency.

View full abstract