電気学会電子・情報・システム部門システム技術委員会では，過去6年にわたり，機械学習技術に関連した延べ5つの調査専門委員会・協同研究委員会を設置・運営をしている。これらの委員会における研究活動を通して，最新の機械学習技術を基礎とする知能システム技術の議論を深め，

抄録全体を表示

For the automatic visual recognition, the semantic gap is the long lasting problem. Recently, by using the internet and the crowd sourcing services, the high quality annotated image datasets have been developed. To maximally utilize the high quality datasets, the strong computational power, and the efficient machine learning methods, the visual recognition system is showing signs of overcoming the semantic gap. In this paper, we overview and explain the recent development of the machine learning and the datasets for the visual recognition.

抄録全体を表示

The lack of data-scientist is big problem in the world. For solving this problem, we develop “Heterogeneous Mixture Learning” technology which automates trial-and-error of data analysis. And also we make many predictive analytics solution using this technology. In this paper, we describe system architecture and analysis method of predictive analytics solution, and also describe problems during operating predictive analytics solution using machine learning.

抄録全体を表示

In this paper, we propose an intelligent security camera system for automated detection of snatching incident. Also, BSAM(Basic Snatching Action Model) is presented to give a definition of the snatching incident. The localization of moving objects in a video stream and human behavior estimation are key techniques for the proposed system. Some motion characteristics are determined from video streams, and using Support Vector Machine, the system automatically classifies the situation of the video streams into criminal or non-criminal scenes. After constructing the classifier, we use test sequences that are continuous video streams of human behavior consisting of several actions in succession. We consider four types of scenarios for the experiments of the snatching incident. The experimental results show that the system can effectively detect criminal scenes at 95.6% accuracy.

抄録全体を表示

Falls are very common among elderly patients in hospitals and nursing homes. This paper presents a method to recognize of getting-up motion which is an early indicator that can be used to help prevent falls. Our method is constructed by autocorrelation of features extracted from edge direction and local intensity gradients. With the focus on the local regions of the image sequence, the features extracted by our proposal method are used as input information for AdaBoost. These features are turned to discriminate between different classes of action. We evaluate our algorithms on 5190 video sequences, containing 2250 getting-up motions and 2940 different actions.

抄録全体を表示

It is known that Improved Penalty Avoiding Rational Policy Making algorithm (IPARP) can learn policies by a reward and a penalty. IPARP aims to identify penalty rules that have a high possibility to receive a penalty. Though IPARP is effective in many cases, it needs many trial-and-error searches due to memory constraints. In this paper, we propose a method called Expected Failure Probability Algorithm (EFPA) to speed it up. In addition, we extend EFPA to multi-agent environments. In multi-agent learning, it is important to avoid concurrent learning problem that occurs when multiple agents learn simultaneously. We also propose a method to avoid the problem and confirm the effectiveness by numerical experiments.

抄録全体を表示

The knowledge concerning an agent's policies consists of two types: the environmental dynamics for defining state transitions around the agent, and the behavior knowledge for solving a given task. However, these two types of information, which are usually combined into state-value or action-value functions, are learned together by conventional reinforcement learning. If they are separated and learned respectively, we might be able to transfer the behavior knowledge to other environments and reuse or modify it. In our previous work, we presented appropriate rules of learning using policy gradients with an objective function, which consists of two types of parameters representing the environmental dynamics and the behavior knowledge, to separate the learning for each type. In the learning framework, state-values were used as reusable parameters corresponding to the behavior knowledge. Instead of state-values, this paper adopts action-values as parameters in the objective function of a policy and presents learning rules by the policy gradient method for each of the separated knowledge. Simulation results on a pursuit problem showed that such parameters can also be transferred and reused more effectively than the unseparated knowledge.

抄録全体を表示

This paper presents a Markov network based multi-objective estimation distribution of algorithm (MMEDA) to solve the resource constrained scheduling problem (RCSP), which hybrid a constraint handling by Markov network based EDA and multi-objective optimization by enforced EDA. Firstly, in order to increase the searching performance while keeping the diversity of Pareto solutions, two kinds of fitness assignment functions are integrated within a novel paradigm. Secondly, Markov network, as an undirected graph model, is adopted to model interrelation between variables with constraints. Thirdly, an enforced EDA with mutation operation is proposed to handle the scheduling. Fourthly, a problem-specific local search for RCSP is applied to improve searching performance. Experiments are conducted on multi-mode resource constrained scheduling problem (MRCPSP) which is an extended RCSP including multi-mode resource constraints. The results of the proposed method highly outperformed conventional meta-heuristic based scheduling methods.

抄録全体を表示

In this paper, the author reports a study into usefulness of polysemy in the 6-multiplexer problem. The study is based on integer linear programming models for the problem of obtaining optimal classifiers for that multiplexer problem. Two classifier designs are considered each of which determines outputs according to inputs and classifiers. One design is typical. In this design, one classifier has one action, and one classifier supports only one output. In the other design with polysemy, one classifier has votes for all actions, and one classifier may support multiple outputs. In both designs, majority voting by matching classifiers determines the output for the corresponding input. Integer linear programming models are developed for some problem settings which differ in the classifier design, the number of classifiers, and the usage of the default rule which does not care the input. Solving those models display that if the number of classifiers is 4 and the default rule is not permitted, the design with polysemy yields more effective classifier sets than the other design.

抄録全体を表示

The research field of human like agents that are often represented by an animation character is becoming increasingly active in recent years. As the motion of such agents influences the users' impression, it is easy to expect that the ability of the human like agent to make appropriate gestures could improve the understandability of the utterance contents. The load of the content creator, however, increases if he/she needs to determine when and what gestures the agent should make. This paper attempts to estimate the appropriate gestures for a given utterance text using conditional random fields (CRF), which can be used to reduce the effort spent by contents creators. We create the dataset consisting of the utterance text and the corresponding gesture labels from the educational movie contents and construct a gesture-labeling model using CRF in a supervised learning manner. The estimation performance of appearing the gestures is evaluated and compared with the simple existing model. Especially, we focus on the metaphoric gesture, often representing an abstract concept. This is because the metaphoric gesture is expected to facilitate the users' understanding of the abstract concepts. We empirically confirmed that the proposed model can distinctly estimate the metaphoric and other gestures.

抄録全体を表示

This paper addresses a generation method on case data for DEA(Data Envelopment Analysis) study using “agents”. Books include some useful business scenario cases and data for a learner to try analysis based on his/her knowledge, however, they are in a sense limited examples to understand deeply the way of DEA. Since generation of various and appropriate cases takes much time and cost, there exists strong need to generate such cases automatically. We propose agent-based software framework to provide various example cases of business scenario and data, which includes self-validated functionalities. Decision tree analysis and evolutionary computational methods are used to construct our software framework, and experimental results show its effectiveness.

抄録全体を表示

In social media many users send personal messages depending on their environments and such messages are used as outputs from a sensor system observing the real world. But the social media is quite different from a general sensor system because users regarded as sensors make messages based on various judgement criteria and the criteria is not controllable. In this study we assume that the judgement criteria occurs according to their belonging communities. So we try to extract messages emphasizing the difference of communities. In this paper we proposed a group specific text discovery method using abnormal detection. We use Twitter as messages generated by social media users. Because the tweets include description of events and tweet generated location, we can extract characteristic tweets based on their generated location. In an evaluation experiment we used tweets related to heavy snow in Yamanashi and found some messages describing local information comparing with tweets except Yanamashi.

抄録全体を表示

The effectiveness of reinforcement learning for the real robots such as crawling robot and six-legged robot was demonstrated in our early studies. However, the acquired movements of those robots were the repetition of one step move and wait for observation at every time step. In other words, the acquired patterns were not smooth forward movements. In this paper, we realize that the multi-legged robots such as six-legged robot and four-legged robot acquire the smooth movements to the target direction and to reach the target area. In order to realize these tasks, we apply reinforcement learning to the multi-legged robots in which Central Pattern Generators (CPG) are implemented. CPG parameters to generate efficient walking patterns are optimized by using reinforcement learning. Through the experiments using the real multi-legged robots, we confirm the effectiveness of our method using CPG and reinforcement learning.

抄録全体を表示

In recent years, along with the popularization of SNS, the incidents, which are called flaming, that the number of negative comments surges are on the increase. This becomes a problem for companies because flamings hurt companies' reputation. In order to minimalize the damage of reputation, we propose the method that detects flamings by estimating the sentiment polarities of SNS comments. Because of the unique SNS characteristics such as repetition of same comments, the polarities of words are sometimes wrongly estimated. To alleviate this problem, transfer learning is introduced. In this research, the sentiment polarities of words are trained in every domain. This will enable to extract the words that are domain-specific and dictate the polarity of comments. These words are occurred in retweets. Transfer learning is implemented to non-extracted words by averaging the occurrence probabilities in other domains. These processes keep the polarities of important words that dictate the polarity of comments and modify the wrongly estimated polarities of words. The experimental results show that the proposed method improves the performance of estimating the sentiment polarity of comments. Moreover, flamings can be detected without missing by monitoring time course of the number of negative comments.

抄録全体を表示

Since 1970s, linear models such as autoregressive (AR), moving average (MA), autoregressive integrated moving average (ARIMA), etc. have been popular for time series data analyze and prediction. Meanwhile, artificial neural networks (ANNs), inspired by connectionism bio-informatics, have been showing their powerful abilities of function approximation, pattern recognition, dimensionality reduction, and so on since 1980s. Recently, deep belief nets (DBNs) which use multiple restricted Boltzmann machines (RBMs) and multi-layered perceptron (MLP) are proposed as time series predictors. In this study, a hybrid prediction method using DBNs and ARIMA is proposed. The effectiveness of the proposed method was confirmed by the experiments using CATS benchmark data and chaotic time series data.

抄録全体を表示

As the Internet is widespread and there are many online shops in the Internet, many persons buy products in the online shops. Customer's behavior in the online shops is a sequence of customer driven activities intrinsically because his/her movement in an online shop occurs according to only his/her decision. Hence, to achieve satisfactory purchase experiments it is important how the shop supports them. Online shops have to predict customer's intents correctly to support them effectively. One of information resources the shops can use is an access log including information on customer's movement in the online shop. If they are histories of customer's behaviors in online shops and the behaviors depend on customer's intents, we can extract knowledge on them from the access logs. Speaking concretely, we can predict customers' intents from the access logs since their internal intents affect their activities. We can realized more appropriate recommendation service by changing recommendation strategy depending on customer's intents. In this paper, we propose a method to predict customer's intents from access logs in a real online shop. We adopt a Topic Tracking Model (TTM) to analyze the access logs.

抄録全体を表示

“Discord” is useful method of detecting anomaly from time series data such as equipment sensor data. The discord method detects anomaly by calculating brute-force nearest neighbor distance between two subsequence of the time series. Hence there is a problem of increasing detection time with the increase in data length. In this paper, we propose an algorithm that, at first, clusters subsequences of the input time series and selects centroid from each clusters (called as “exemplar subsequence”), then detect anomaly by calculating nearest neighbor distance between input subsequence and exemplar subsequence. We also show that our proposed method is more faster detection speed than the discord method and have equivalent detection ability as the discord method by an experiment with two time series data.

抄録全体を表示

The authors have proposed the signal decomposition technique as one of the powerful solution to mitigate the large PAPR to be addressed in OFDM transmitters especially on mobile terminals. In order to enhance the receiver SNR, the simple noise elimination techniques working together with the signal decomposition technique have also been proposed, that eliminates the noise added on the decomposed constant amplitude on-off-signals taking advantage of the knowledge of their constant amplitude at the receiver. In this paper, we discuss the parameter design issue of the proposed techniques and their optimization. Then, demonstrate the PAPR, the PAE and also the BER performances operating on the optimized parameters. It is confirmed that the proposed signal decomposition technique improves the PAPR by 4dB and doubles the PAE at the CCDF (Complementary Cumulative Distribution Function) of 1%. It is also confirmed that the proposed noise elimination technique improves the receiver SNR by 3dB at the BER of 10^-3, which is nearly equal to that of conventional OFDM, under the conditions that the decomposed signals are transmitted over the independent AWGN channels. Furthermore, it is demonstrated that the proposed techniques work properly when the decomposed signals are transmitted over 2×2 MIMO.

抄録全体を表示

We report experimental validation of the existence of so-called the sleep rebound phenomenon against each sensory stimulus using a driving simulator and the electroencephalograph for about 105 subjects with evaluation of an arousal index (α+β)/(δ+θ) in the electroencephalogram (EEG) measurements after each driving operation. We found that methods using the perfume presentation, the alert presentation, the vibration and chewing gum were resulted in the sleep rebound phenomenon, whilst a proposed magnetic stimulation showed an arousal retention effect without the sleep rebound. Mechanisms of the arousal effect of the magnetic stimulation were discussed with measurements of electric conductivity versus water temperature characteristics of pure water.

抄録全体を表示

This paper presents a new mathematical model for robust meter placement against false data injection attacks on state estimation in power grids. It has been recently reported that the false data injection attacks can introduce arbitrary errors into the output of the state estimation without being detected. While one of the most promising strategies against attacks is to protect meters, it may not be feasible to protect all meters due to some reasons such as budgetary constraints. It is important to select a subset of meters to be protected so that, even if any k meters are compromised by attackers, we can detect the attacks with the rest of them. Several algebraic approaches have been proposed to determine the minimum required meter set to detect the attacks even if any k meters fail. However, these approaches have a common problem in that the attacks on both of existing and additional meters are not considered. In this paper, to remedy this problem, we propose a new model for robust meter placement against false data injection attacks. Through the numerical experiments with the IEEE test systems, we investigate the total computation time required to obtain the optimal meter placement and the number of required additional meters. As the results, for the case of k=2 and 3, we succeed in obtaining optimal meter placements up to the 300-bus and the 57-bus systems within one hour of computation time, respectively.

抄録全体を表示

Nonlinguistic information such as facial expression, gesture, and tone of voice plays an essential rule in communication between humans. However, most human-machine interfaces are not designed to utilize nonlinguistic information. In order to develop a natural human-machine interface, this paper presents a system for discriminating positive-negative attitudes from nonlinguistic information. Our system utilizes prosody information and natural facial expression observed in a dialogue. For example, the questioner asks a subject to do something, e.g. “Won't you go shopping with me?”. And then, the subject shows an ambiguous response, e.g. “Umm, shopping...”. The proposed system analyzes the prosody information and facial expression obtained from the subject. The system combines optical flows and prosody information to increase the accuracy in the discrimination. Evaluation experiments using 678 videos obtained from 23 subjects have shown that our system achieved 81.4% accuracy, which is 4.8% higher in the case when only optical flows were used as training data and 13.8% higher in the case when only prosody information was used as training data.

抄録全体を表示

Onomatopoeia is a powerful means to convey characteristics of sounds and also important for designing sounds. We conducted an experiment wherein 493 people, aged 7-81 years, and 376 kindergarteners, aged 3-6 years, were requested to imitate pure sounds of 62.5 Hz, 500 Hz, and 4 kHz with onomatopoeic voices. We analysed differences in trends of onomatopoeic expressions based on gender and age. Various onomatopoeic voices were produced by the subjects, many of which were those starting with consonants “P” or “B” followed by long vowels. There were significant gender differences in their starting consonants at 500 Hz and 4 kHz, whereas age differences were remarkable at 62.5 Hz and 500 Hz. A significant difference was observed in both gender and age at 500 Hz for long vowels. The onomatopoeic voices indicated by the kindergarteners had a lot of variations, among which their consonants in particular were found to be significantly different from that of adults at all frequencies. Noticeable gaps were found between three-year-old children and adults in vowels, but the gaps in vowels of the children became closer to those of adults as they aged.

抄録全体を表示

The portfolio optimization problem involves decisions pertaining to the investment target and proportion of investment in a large number of assets in order to minimize risk and maximize returns. In recent years, metaheuristics methods have been actively applied to portfolio optimization. Under portfolio optimization, the portfolio is optimized for a fixed period of time so that its performance during that period is excellent. However, the optimized portfolio may not be able to sustain that performance later. Therefore, there is a need for recombining assets and changing the proportion of asset allocation by means of rebalancing. The rebalancing has to be done at an appropriate time. In this paper, we propose a technique for dynamic rebalancing of a portfolio at an appropriate time by applying instance-based policy optimization, with a consideration of market conditions changes.

抄録全体を表示

We propose a method for modelling navigation structure of web application using state machine diagrams. We assume that the navigation structure is described by a page flow diagram. The navigation structure is modelled as state machine diagrams representing page navigations, authentication and user operations. A model of page navigations is obtained by representing web pages and movement of them as states and transitions, and by representing form controls in web pages as substates. A model of authentication is obtained by representing a condition of authorization as states which changes by events corresponding to user operations. User operations are represented as actions in transitions. We applied the proposed method to an example web application described by a page flow diagram and modelled it as state machine diagrams. We demonstrated automatic verification and test generation for the obtained state machine diagrams using SAL tools. We showed that errors in the page flow diagram can be detected successfully and the generated test case can cover all navigations. Additionally, we applied the method to a practical example and showed that verification and test generation by SAL can be carried out in a small amount of time.

抄録全体を表示

In this letter, a two-tone signal design method is introduced to achieve high-efficient linear power amplifier (PA). The proposed method, which embraces harmonic tuning, can be employed to design a PA with high efficiency and low nonlinear distortion. To verify the proposed method, a high-efficient linear PA is implemented using GaN HEMT. Experimental results using a 5 MHz two-tone signal show that a drain efficiency (DE) of 55.6% and a carrier-to-intermodulation ratio (C/I) of 30 dB was achieved at 3.47 GHz. The PA shows a DE of 46.6% when stimulated by a 20-MHz long-term evolution (LTE) signal with 8.2-dB PAPR.

抄録全体を表示