In this paper, we propose a method for acquiring shape models of 3D objects from range data of objects in a class. Since objects in a class have various kinds of structures, a shape model is generated for every structure. First, a range image is segmented into parts based on the curvature of a surface, and the part is approximated as superquadrics. Superquadrics are parametric representation for a 3D shape; their parameters can be criterion for similarity between 3D shapes. Next, integration of parts is required because the segmentation based on curvature makes excessive parts. For this purpose, ``Part agent'' is assigned to the part, and integrates parts by interaction based on a fitting error. From the results, input objects are clustered into groups. This clustering method is based on interaction by ``Object agent'', which is assigned to the object. First, the object agents in charge of same structured objects make a group. An object agent evaluates groups based on an evaluation function, then move to the group with highest evaluation. The evaluation function is based on the fitting error in the case of belonging to the group, the distance from the average shape of the group, the scale of the group, and the number of parts. A clustering result is obtained when the agents do not move. Since each cluster includes only same structured objects, the object in the cluster can be represented as the same number of parameters. Therefore, a shape model is generated by manipulating parameters. Finally, the experimental results show that the proposed method can acquired valid shape models and can improve segmentation results.
E-Commerce sites have introduced new business models for effective and efficient commerce. If we introduce software agent technologies into e-commerce systems, we can further enhance the intelligence of user support. In this paper, we propose a new cooperation mechanism among seller agents based on exchanging their goods in E-GarageSale, our P2P-based distributed electronic marketplace. In E-GarageSale, buyer agents can purchase multiple goods based on discount prices. However, there is a case that seller agent does not have enough goods for trading with several buyer agents at a time. Therefore, seller agents cooperatively negotiate by using an exchanging mechanism for selling goods effectively. By using our mechanism, all seller and buyer agents can increase their utility. Our experiments show that an exchanging mechanism enables seller agents to sell goods in stock effectively.
In this paper, we present a method to deal with the quantitative belief change based on linear algebra. Since the epistemic state of an agent is represented by a set of the subjective probability which she conceived for each possible world, we can regard this epistemic state as a point of vector space spanned by the basis of possible worlds. The knowledge which causes the belief change is treated as a matrix on this vector space. The observation of the new fact about the current world is characterized as a projection matrix. On the other hand, the knowledge about the world change is represented by a stochastic matrix. In this framework, we present a unified method of belief change both for propositional and probabilistic knowledge so that the logical or probabilistic reasoning in the process of the belief revision and update is reduced to the matrix calculation.
Learning from cluster examples (LCE) is a composite task of two common classification tasks: learning from examples and clustering. Learning from cluster examples involves an attempt to acquire a rule that can be used to partition an unseen object set from given examples. A general method for learning such partitioning rules is useful in any situation where explicit algorithms for deriving partitions are hard to formalize, while individual examples of correct partitions are easy to specify. In this paper, to improve estimation accuracy of LCE, we employ attributes of clusters and propose a method that can handle this type of attributes. We show improvements of performance by applying this method to artificial data.
In this paper, we describe an improvement of a calculation procedure of logic programs. The procedure proposed before is the combination of a replacement procedure of logical formulae and a transformation procedure of equations to disjunctive normal form, and it can calculate logical consequences of the completion of any given first-order logic program (FLP), which is equivalent to the FLP in two-valued logic, soundly and completely in three-valued logic. The new procedure is also the combination of them, but the transformation procedure is improved to be able to calculate two-valued logical consequences of the FLP more than the old one. We prove that it can calculate logical consequences of a completed program, which is not equivalent to the completion of the FLP, soundly and completely in three-valued logic.
In this paper, we present an active audition system which is implemented on the humanoid robot "SIG the humanoid". The audition system for highly intelligent humanoids localizes sound sources and recognizes auditory events in the auditory scene. Active audition reported in this paper enables SIG to track sources by integrating audition, vision, and motor movements. Given the multiple sound sources in the auditory scene, SIG actively moves its head to improve localization by aligning microphones orthogonal to the sound source and by capturing the possible sound sources by vision. However, such an active head movement inevitably creates motor noises.The system adaptively cancels motor noises using motor control signals and the cover acoustics. The experimental result demonstrates that active audition by integration of audition, vision, and motor control attains sound source tracking in variety of conditions.onditions.
``The rhetoric of the film'' means the combination of the film techniques based on a film director's purpose. A system of computational film generation must include the rhetorical aspects. In this paper, we discuss using computational film creation as a type of rhetoric in our strategy for triggering a viewer's cognitive process. We especially consider the relationships between identical and different elements in shots of a film. Then, a system for film generation based on the classification of the rhetoric is proposed. The system generates various types of films according to a rhetorical strategy for intensity of cognitive effects. The generated films can cause intensive cognitive effects in a viewer. For example, one generated film using location as ``different elements'' can make viewers cause ``resetting of a viewpoint'' and ``affect which is not arose by understanding the story, but by the audiovisual situation.'' This research corresponds to the aspect of non-story processing in a whole narrative generation mechanism which includes both aspects of story and non-story processing.
This paper presents a developmental learning model for joint attention between a robot and a human caregiver. The basic idea of the proposed model comes from the insight of the cognitive developmental science that the development can help the task learning. The model consists of a learning mechanism based on evaluation and two kinds of developmental mechanisms: a robot's development and a caregiver's one. The former means that the sensing and the actuating capabilities of the robot change from immaturity to maturity. On the other hand, the latter is defined as a process that the caregiver changes the task from easy situation to difficult one. These two developments are triggered by the learning progress. The experimental results show that the proposed model can accelerate the learning of joint attention owing to the caregiver's development. Furthermore, it is observed that the robot's development can improve the final task performance by reducing the internal representation in the learned neural network. The mechanisms that bring these effects to the learning are analyzed in line with the cognitive developmental science.