The semantic similarity (or distance) between words is one of the basic knowledge in Natural Language Processing. There have been several previous studies on measuring the similarity (or distance) based on word vectors in a multi-dimensional space. In those studies, high dimensional feature vectors of words are made from words' cooccurrence in a corpus or from reference relation in a dictionary, and then the word vectors are calculated from the feature vectors through the method like principal component analysis. This paper proposes a new placement method of nouns into a multi-dimensional space based on words' cooccurrence in a corpus. The proposed method doesn't use the high dimensional feature vectors of words, but is based on the idea that ``vectors corresponding to nouns which cooccur with a word w in a relation f constitute a group in the multi-dimensional space''. Although the whole meaning of nouns isn't reflected in the word vectors obtained by the pro posed method, the semantic similarity (or distance) between nouns defined with the word vectors is proper for an example-based disambiguation method.
This study argues how human infants acquire the ability of joint attention through interactions with their caregivers from a viewpoint of cognitive developmental robotics.
In this paper, a mechanism by which a robot acquires sensorimotor coordination for joint attention through bootstrap learning is described.
Bootstrap learning is a process by which a learner acquires higher capabilities through interactions with its environment based on embedded lower capabilities even if the learner does not receive any external evaluation nor the environment is controlled.
The proposed mechanism for bootstrap learning of joint attention consists of the robot's embedded mechanisms: visual attention and learning with self-evaluation.
The former is to find and attend to a salient object in the field of the robot's view, and the latter is to evaluate the success of visual attention, not joint attention, and then to learn the sensorimotor coordination.
Since the object which the robot looks at based on visual attention does not always correspond to the object which the caregiver is looking at in an environment including multiple objects, the robot may have incorrect learning situations for joint attention as well as correct ones.
However, the robot is expected to statistically lose the learning data of the incorrect ones as outliers because of its weaker correlation between the sensor input and the motor output than that of the correct ones, and consequently to acquire appropriate sensorimotor coordination for joint attention even if the caregiver does not provide any task evaluation to the robot.
The experimental results show the validity of the proposed mechanism.
It is suggested that the proposed mechanism could explain the developmental mechanism of infants' joint attention because the learning process of the robot's joint attention can be regarded as equivalent to the developmental process of infants' one.
Recently, data mining is remarkable as a practical solution for huge accumulated data. The classification, the goal of which is that a new data is classified into one of given groups, is one of the most generally used data mining techniques. In this paper, we discuss advantages of Memory-Based Reasoning (MBR), one of classification methods, and point out some problems to use it practically. To solve them, we propose a MBR applicable to business problems, with self-determination of proper number of neighbors, proper feature weights, normalized distance metric between categorical values, high accuracy despite dependent features, and high speed prediction. We experimentally compare our MBR with usual MBR and C5.0, one of the most popular classification methods. We also discuss the fitness of our MBR to business problems, through an application study of our MBR to the financial credit management.
This paper presents the Real-coded Genetic Algorithms(RCGA) which can treat with high-dimensional ill-scaled structures, what is called, k-tablet structure. The k-tablet structure is the landscape that the scale of the fitness function is different between the k-dimensional subspace and the orthogonal (n-k)-dimensional subspace. The search speed of traditional RCGAs degrades when high-dimensional k-tablet structures are included in the landscape of fitness function.
In this structure, offspring generated by crossovers is likely to spread wider region than the region where the parental population covers. This phenomenon causes the stagnation of the search. To resolve this problem, we propose a new crossover LUNDX-m, which uses only m-dimensional latent variables. The effectiveness of the proposal method is tested with several benchmark functions including k-tablet structures and we show that our proposal method performs better than traditional crossovers especially when the dimensionality n is higher than 100.
Most of real world problems contain complex and various constraints, and the penalty depending on the degree of violation is often used to handle them. However, two objectives, to reduce the violation and to optimize the primary value, are inherently oppositive. Therefore, using additive penalty method (APM) often leads the fatal compromise to a solution with bad primary objective value in return for no violation. In this paper, we employ separated constraint satisfaction (SCS), to deal these two objectives independently like as a multi-objective optimization. The efficiency of SCS is shown on a simple benchmark and the sewerage system control problem.
This paper discusses a problem of human-machine interaction when spoken word to object reference ambiguity occurs. We study joint activity of several agents in which a remote robot finds an object while communicating with the user over a voice-only channel. We focus on the problem in which the robot disambiguates the reference of the uttered word or phrase to the target object. For example, the utterance of the word ``cup'' may refer to a ``teacup'', a ``coffee cup'', or even a ``glass'' for different users in some situations. This reference (hereafter, ``object reference'') is user and situation dependent. We conducted two experiments. The first experiment including 12 subjects confirmed that the user model of object references is significant. In the second experiment conducted on 20 subjects, we show the model reference sensitivity to the situation. In addition to the ambiguity of the object reference, the actual system must cope with two sources of uncertainty: speech and image recognition. We present the belief network based probabilistic reasoning system to determine the object reference. The resulting system demonstrates that the number of interactions needed to find a common reference is reduced as the user model is refined.
Belief fusion, instead of AGM belief revision, was first proposed to solve the problem of inconsistency, that arised from repetitive application of the operation when agents' knowledge were amalgamated. In the preceding work of Maynard-Reid II and Shoham, the fusion operator is applied to belief states, which is total preorders over possible worlds which is based on the semantics of belief revision. Moreover, they introduced the pedigreed belief state, which represented multiple sources of belief states, ordered by a credibility ranking. However in the theory, all the sources must be totally ordered and thus applicable area is quite restrictive. In this paper, we realize the fusion operator of multiple agents for partially ordered sources. When we consider such a partial ranking over sources, there is no need to restrict that each agent has total preorders over possible worlds. The preferential model, based on the semantics on nonmonotonic reasoning, allows each agent to have strict partial orders over possible worlds. Especially, such an order is called a preferential relation, that prescribes a world is more plausible than the other. Therefore, we introduce an operation which combines multiple preferential relations of agents. In addition, we show that our operation can properly include the ordinary belief fusion.
This paper presents Multi-State Real-Time Bidirectional Search (MSRTBS), a method that improves the efficiency of the heuristic search algorithm for finding approximate solutions. Real-Time A* (RTA*) is a representative heuristic search algorithm for finding approximate solutions. The Multi-State Commitment (MSC) method was introduced into RTA* and dramatically improved the performance in problems such as the N-puzzle. As well, Real-Time Bidirectional Search (RTBS) also improved RTA* by changing a unidirectional search into a bidirectional one. This paper proposed a method that introduces MSC into RTBS. The experimental results showed that compared with RTBS and MSC our proposed method, MSRTBS, improved executed time and solution quality in the N-Puzzle.