This paper proposes an iterative greedy combinatorial auction (CA) mechanism for multi-agent pathfinding (MAPF). MAPF algorithms assign agents non-conflicting paths that minimize a global cost (e.g. the sum of travel costs). CA mechanisms allocate agents items that maximize social welfare. Recent works proposed two mechanisms that coordinate self-interested agents in MAPF: one is called iterative taxation framework (ITF), which is similar to the simultaneous ascending auction, and the other is an iterative CA mechanism, which guarantees to minimize the global cost. However, both methods have drawbacks: the former has no approximation guarantee, and the latter suffers from the fact that winner determination in the CA is computationally infeasible (NP-hard). In this paper, the author propose two iterative greedy CA (IGCA) mechanisms for MAPF that are computationally efficient. The first mechanism provides square of the number of agents approximation. Second mechanism was shown, by numerical simulations, to give a lower cost than ITF.

Transparency in machine learning (ML) agents’ decision making is crucial for achieving AI Safety. However, it is difficult to comprehend agents’ behavior because gaps of perception, mobility, desire, and time scale between humans and ML agents obstruct people to mentalize the agents. In this paper, we propose a model of human’s inference of ML agents’ mental states so as to explain the agents’ behavior from the perspective of humans.

The representation of Constraint Optimization Problems for multiagent systems has been addressed in several studies including Distributed Constraint Optimization. We focus on the asymmetric problem where each objective function is defined as an evaluation of an individual agent. While this class of problems is studied as a multi-objective problem, there are opportunities to investigate various types of solution methods. In this study we investigate the possibility of the framework based on the Lagrangian dual methods. We address a bottleneck problem that minimizes the worst case cost. As the initial study, we experimentally apply and evaluate a formalization of the problem.

In recent years, Social Networking Serveices (SNSs) have been popular among young people. According to the popularity of SNSs, the cyberbullying becomes a social problem. To find the mechanism of cyberbullying, there are several studies using multi-agent simulation which can analyze human relationships. Therefore, in this study, we model SNSs in classroom and analyze the influence of SNSs on classroom friendship. This model is based on the Socion theory. In the model, an agent can communicate other agents using two kinds of networks: a network of classroom and networks of SNSs via a network recognized by each individual. Agents communicate face to face (FTF) in the classroom and do SNS communication in SNSs. In this paper, we conduct simulation of friendship relations considering SNSs, and disccuss the influence of SNSs on classroom relationships from the simulaion results.

We have been working on a conversational dialogue system that raises user satisfaction and draws interests through dialogues, unlike a task-oriented dialogue that answers simple instructions and questions. To raise user satisfaction, systems should precisely respond to a user utterance and continue dialogue with topics that strongly related to the user utterance. However, most conversational systems, which are usually utilize one-turn query-response pairs, are difficult to naturally continue dialogue. In this study, by controlling dialogue topics with multiple robot coordination, we propose a novel conversational system with two-turn query-response pairs that can naturally continue consistent dialogue. Our field experiment at Kyoto city zoo shows that our proposed system achieves very high user satisfaction.

There is a possiblility of change from the best coalition structure formation at the moment to the best coalition structure at the next moment in a real-world coalition formation scenario. In this paper, we present a preliminary approach about constructing a continuous cooperative game simulator, which is focused on the situation where a coalition structure can be changed by adding new conditions of coalition formations. This preliminary simulator takes a list of pair of timing and a MC-nets rule added as a continuous cooperative game. We define a distance between two coalition structure as its costs of changes to model Inertia. We present how a coalition formation simulation can be done with such as inertia.

By improvement of computer’s performance, it has become to enable to simulate the behavior of structure, soil and so on using the nonlinear structural analysis of large and complex models. However, these simulations require lots number of convergent calculation to search a convergent solution. Although there are several convergent methods such as Newton’s method, there is no almighty method that fits all situations. Hence, it is up to the person carrying out the analysis to select one of these methods. Considering this situation, in this paper, we propose a method to choose and combine the appropriate convergent method using the concept of Q-learning from reinforcement learning. First, using simple analysis model, we train an action value table to choose the appropriate convergent method with Q-learning. Then, we carried out an analysis using obtained action value table and show that our method can make the convergence time shorter than conventional ones.

Evacuation planning is an important issue to estimate cost and safety of buildings in the basic design phase for the architectural design. Building safety is evaluated according to evacuation safety verification method, multi agent simulation model and otherwise. However, specialized knowledge about human behavior of the building evacuation is insufficient. In this paper, an evacuation simulation of total evacuation drill in office building using multi agent model is conducted. Moreover, the result of simulation is compared with acquired measurement data of actual evacuation behavior especially for evacuation time and route selections. The route selections of simulation are different from that of actual behavior.

In order to carry out prompt restoration activities after the occurrence of a natural disaster such as an earthquake, it is important to grasp the buildings’ damage situation and damage distribution of the affected area as soon as possible. In this study, the buildings’ damage is evaluated from the image data of building by using Convolutional Neural Network. Furthermore, to grasp the geographical distribution of the damage, a results of the damage evaluated by CNN are plotted on a map.

In urban design, design considering pedestrian flow lines is important. Large scale redevelopment around Nagoya station is planned due to opening of Linear Central Shinkansen 's Nagoya Station. Along with redevelopment, it is necessary to design a flow line that is easy to use for users by considering flow lines around the station. In this research, by collecting actual human flows at major stations in Nagoya city and modeling it, it is utilized for simulation aimed at redevelopment of cities.

We tried to identify the state of members injured by natural disasters by machine learning using the vibration response waveform of a piezoelectric sensor attached a wooden element. We destroyed step by step a wooden wall connected to columns. In each destruction stage, the vibrational characteristic is measured by a piezoelectric sensor. The oscillating source to obtain data models vibrations of natural vibration coming from a wind and so on. We detected the injury of a member by using autoencoder which learned only the waveform of the element which is not injured.

Recently, utilization of regenerated power from the brake operation has drawn attention to help for energy conservation of railway systems. In this paper, we introduce reinforcement learning with actor-critic to acquire the appropriate rules that decide the amount of charge/discharge so that the fluctuation of the SOC (State of Charge) can be suppressed.In the previous researches, the control rules are hand-crafted based on human empirical heuristics that has limitations on suppressing electricity supply-demand dynamics in the railway system. We show some empirical results of both previous research and our proposed one, and find that the generated rules via reinforcement learning show the better performance than the ones by the previous researches.

Realization of a sustainable low-carbon society is required globally, and energy consumption reduction derived from fossil fuels has been emphasized. In the transportation sector, which accounts for a large percentage of energy consumption, EVs with high environmental performance has been popularized and it is important to use renewable energy for its charging effectively. In this research, we proposed a coupled simulation model that can represent interactions between transport and electric power systems. Since the charging event is coupled to the electric power system mechanism, the spread of EVs will affect not only the road transport network but also the electric power system. As a result of the numerical experiments, it was implied that the concentration of low-output charge after returning home might cause a voltage drop in the distribution system. This phenomenon became more prominent as the penetration rate of EV increased, and the possibility that it deviates from the legal range of the reference voltage is shown in some scenarios.

Trip demand prediction plays a crucial role in bike-sharing systems. Predicting trip demand is a highly challenging problem because it is influenced by multiple factors, such as periodic changes, correlation between stations, weather and types of users. Although several recent studies successfully address some of these factors, no framework exists that can consider all of them simultaneously. To this end, we develop a novel form of the point process that jointly incorporates all the above factors to predict trip demand, i.e., predicting the number of pick-up and drop-off events in the future and when over-demand is likely to occur. Our extensive experiments on real-world bike sharing systems demonstrate the superiority of our trip demand prediction method over five existing methods.

In this paper, we explore the utility of customer behavior analysis of un-structured video data for improving sales forecasting, which is an important task for supply chain management in retail store. Specifically, we develop video-based customer behavior analysis system for monitoring and analyzing customer's shopping behavior, then extract the information about how the customers interact with the stores and products, and finally design a framework to incorporate the video-based analysis result into sales forecasting. To the best of our knowledge, this is the first work that introduces the customer behavior analysis of monitoring video data to sales forecasting task. In order to validate our observation, we conducted a series of experiments in a physical retail store and demonstrated that integrating video-based customer behavior analysis into a conventional sale forecasting model results in a performance improvement.

Special days such as public holidays have significant impact on time-varying customer’s demand. In this paper, we propose a new method to handle the special days' effects by embedding techniques using neural network. We evaluate the usefulness of our method in the real call center data set and demonstrate that embedded features generated by our method provide substantial performance gains in call arrival prediction. In addition, we visualize the embedding features of special day and its neighborhood and further understand their relationship.

We propose to apply Recursive Bayesian Computation to image estimation of SPoD-ExPAN microscopy.The method does not need derivatives of the optimality measure and is supposed to derive images globally better than those of gradient dissent based approaches.In this paper, we present an implementation of the Recursive Bayesian Computation by using Kernel density estimation.Moreover, we introduce regularization to the estimation, and experimentally compare its performance with the case without the regularization.

Extracting objects such as clouds in weather images in the spatiotemporal data and tracing the changes of the shape, the position, the interaction between objects and events such as generation, extinction are important tasks in spatio-temporal data mining. The method of object extraction and tracking using mixture of multivariate normal distribution applying greedy EM algorithm is proposed. Experiments using artificial data showed effectiveness of the proposed method.

Multimodal data including images, sounds, texts is accumulated on the Internet. We can expect general-purpose data representation to perform tasks such as data discrimination, generation, and retrieval on various modalities datasets. The key idea for acquiring the representation is embedding a point from a data space of each modality in a point of common space. However, if data is embedded in a point, it becomes difficult to interpret the ambiguity of the data's meaning and the inclusive relation among the data. Of course, representation of data point does not necessarily need to be a point. In this study, we embed image and text into a normal distribution in a common space. This improves the performance of image retrieval.

Popular songs change with the time and the time changes with popular songs, a proverb says. Hit songs therefore tend to reflect the state of people’s morality that exists when they are written. With the Moral Foundations Dictionary, we quantified morality in Japanese lyrics from musicians who ranked in the Oricon hit chart between 2004 and 2016. The results show that Japanese hit songs include many words related to “harm” and “purity.” In addition, we found a positive correlation in the usage of morality- and immorality-related words.

Social supports such as giving empathy, affection, and respect increase children’s well-being and reduce stress. We aim to drive social support for bullied children through the Internet (online social support). In this study, we focus on bullying in real life (offline) and support through online communication. For this purpose, we investigate how online social supports affect bullied children and which conditions ensure the supports have positive effects by analyzing online communication data on an avatar chat service. We found that bullied-experience talks in the virtual world with a few friends have a positive effect on bullied children. The talks tended to include bullied children’s self-disclosure. Therefore, this finding suggests that self-disclosure by bullied children is very important for supporting them.

Information is transmitted through websites, and the immediate reaction to various information is required. Hence, the efforts for readers to select information themselves have increased, which leads to the further improve- ment of recommendation services that can reduce such burdens. On the other hand, it is pointed out that filter bubbles that only provide biased information to users are generated due to the redundant recommendation. In this research, we analyzed behavioral changes prior to recommendation by clustering, and showed that behavior changes after average number of browsing number and article type change during the period.

People often belong to multiple communities and accommodate themselves to majority opinion in each community. Thus, they sometimes express different and occasionally inconsistent opinions in the individual communities. When the inconsistency in opinion is uncovered, she/he may be criticized, and then, quit expressing opinions. Recently inconsistency in opinion is easily uncovered because of the growth of communication tools, e.g. SNS. Therefore, we investigate the mechanism of quitting expressing opinion because of the criticism of the inconsistency in opinions in social media. We modeled multiple communities as multiplex networks and investigate how opinions are formed and why users quit expressing opinions due to uncovered inconsistency. For this purpose, we first extended a conventional model whose opinion formation mechanism is based on the bounded confidence model. Then we also considered different opinion formation mechanisms and examined opinion formation in different types of multiplex complex networks, in order to adjust more realistic situations.

In recent years, youth low interest about political interest has become a problem in Japan. This problem has has been discussed by questionnaires, but surveys based on actual behavior are rarely carried. In this research, using the user action logs in news curation service, analyze how the viewing trend of news articles on politics varies depending on ages.

We investigate the network structure of global supply chain using a dataset that contains information on business partners for about 500,000 firms worldwide. First, we show that this network has scale-free topology and that the shortest path length is around six. Second, we show through community structure analysis that the firms comprise a community with those firms that belong to the same industry but different home countries, indicating the globalization of firms’ production activities. Finally, we discuss what such production globalization implies for the proliferation of conflict minerals (i.e., minerals extracted from conflict zones and sold to firms in other countries to perpetuate fighting) through global supply chain. We show that several bridge firms between some specific industries and countries play an important role in the global proliferation of conflict minerals.

This paper proposes benchmark functions for cooperative coevolution (CC), which dynamically divides a target problem into smaller subproblems and solves in parallel.The proposed benchmark functions allow adjusting dependency strength between variables by changing the number of linkages and linkage types.Experimental results showed the in uence of dynamic decomposition change frequency to CC's optimization performance.

In this talk, we introduce an idea for incorporating information encoded in a knowledge graph to a latent variable model (LVM). We propose an extension of an LVM by two-view modeling, where the parameters and the latent variables of the LVM are shared between the original LVM and a probabilistic model for knowledge graph embedding. We specifically introduce how to incorporate a knowledge graph into probabilistic principal component analysis and show preliminary experimental results.

In this study, we consider the covariates shift caused by the action on the data problem. Also, consider the method of bias correction using the density ratio estimation.In the numerical experiments, the action actually cause unsteady / extrapolation problem of distribution. It shows that have a significant bias to the expected value evaluation.Further illustrates that the method using the density ratio even in the case of changing the data size and dimensions are effective.

Linearly solvable Markov decision process (L-MDP) is an essential subclass of MDP to find a better policy efficiently. We first develop a novel batch reinforcement learning algorithm for L-MDP in discretized action space. The algorithm simultaneously learns a state value function and a predictor of state values at next step by using pre-collected data. We evaluate our method on traffic signal control domain in a single intersection with the traffic simulator SUMO. Our experiment demonstrates that our method finds the policy on the domain efficiently.

We present Cappuccino, a bootstrapping method for structural data, which extracts target pairs of words by treating structures in data as patterns that connect pairs of words. The task we tackle in this paper is relation extraction from structural data such as presentation slides and HTML files. In order to exploiting the structures for bootstrapping, we convert structural data to graphs, and apply frequent subgraph mining to the graphs to calculate reliabilities of word pairs and patterns. Experimental results of a relation extraction task show that Cappuccino outperforms a previous method.

Non-negative matrix factorization (NMF) is widely used for various problems, such as item recommendation of shopping, image recognition and bioinformatics. However, when the data are sparse, and the row and columns of the items have no association with others, these items tend to be independent resulting poor linking. Here we investigated a method to compensate such too sparse data, by adding information of hierarchical relationships among the items to the matrix analyzed. We show how the additional information helps to make desired new clusters. In addition, too much augmentation of class information makes many items into the same clusters, the situation in which is not desired. Some ideas to avoid excessive compensation are discussed.

In the sampling based direct policy search in reinforcement learning, higher dimensional decision variables causes the deterioration of optimal value and the slowing down of the learning speed. We clarified that the variance of the sampling probability distribution affects both for the optimal value and the learning speed. Especially, there exists the tradeoff between the optimal value and the learning speed. In this paper, we propose two trick to improve the learning speed without deteriorating the optimal value. First trick is to employ the small variance sampling distribution for improving the optimal value; It causes slower convergence as a side effect. As the second trick, we employed the dimensionality reduction of the decision variable for improving the learning speed.

In apprications of machine learning, we often need to generate some models for multiple similar cases. In our study, we propose an efficient learning method exploiting the idea of the curriculum learning. Our idea is that we evaluate the effectiveness of a curriculum learning for one problem, then we use the learned curriculum for learning other cases. We apply our proposal method to two small problems, addition and subtraction of integers, using Sequence-to-Sequence model. Our experimental results show that the method works efficiently for some problem settings.

Reinforcement learning aims to find a policy which maximizes long term future reward by interacting with unknown environment through trial and error. In this study, we propose an objective correction method for entropy regularized Markov decision process. After deriving a policy gradient under the regularization by the entropy and relative entropy, we propose an on-policy objective correction method for off-policy policy improvement under entropy regularization.

In this paper, We evaluated the time-series gradient boosting decision tree method using benchmark data. Our time-series gradient boosting tree has weak learners with time-series and cross-sectional attribute in its internal node, and split examples based on dissimilarity between a pair of time-series or impurity between a pair of cross-sectional attributes.It has been empirically observed that the method induces accurate and comprehensive decision trees in time-series classification, which has gaining increasing attention due to its importance in various real-world applications.

Graphical processing units (GPUs) have become ubiquitous because they offer the ability to perform cost and energy efficient massively parallel computation. We investigate forward search classical planning on GPUs based on Monte-Carlo Random Walk(MRW). We first propose Batch MRW (BMRW), a generalization of MRW which performs random walks starting with many seed states, in contrast to traditional MRW which used a single seed state. We evaluate a sequential implementation of BMRW on a single CPU core and show that a sequential, satisficing planner based BMRW performs comparably with Arvand, the previous MRW-based planner. Then, we propose BMRW_G, which uses a GPU to perform random walks. We show that BMRW_G achieves significant speedup compared to BMRW and achieves competitive performance on a number of IPC benchmark domains. This is an extended abstract of an ICAPS2018 paper [Kuroiwa 18].

Sequential Monte Carlo (SMC) is a typical sampling method that can be sampled in order from a sequential probabilistic model. However, due to degeneration of the sample, SMC may produce samples with low likelihood with a small number of particles. In our study, we focus on the fact that the same resampling targets of SMC for each sample cause degenerating samples. We want to relax this constraint, but analytically deriving asymmetric sequential resampling targets is difficult. Therefore, we expand resampling strategy of SMC asymmetrically by learning the sequential resampling target from the target of the whole series approximated to the lower bound. By this, by learning to resample, it is expected that accurate estimation of latent variable can be realized with the same particle number as SMC.

I propose a prediction model for end-of-query detection. It is a common practice that a chatbot system responds an answer when they get a single utterance from the user. A problem arises when a human breaks their query into multiple utterances and a chatbot system can not respond appropriately, leading to lower user experience. In this paper, I propose a prediction method based on LSTM for end-of-query detection.

We will demonstrate a counterargument generation system in debating, which aims to utilize newswire articles for decision-making support. Users can specify a claim such as “We should legalize casinos because they promote economy.” and then the system outputs counterargument scripts against the claim.