In this study, the ability estimation in the job interview scene was carried out. In the field of communication between human and human, the communication ability estimation using multimodal information shows high accuracy. It is considered that the same ability estimation is possible in the dialogue between the interviewee and the interviewer. In this study, we developed a model to estimate the evaluation of interviewees by using multimodal information such as speech features, prosodic features, motion features, and head features such as head movement and gaze. In the evaluation of the interviewee, the following were used: Social basic ability determined by the Ministry of Economy, Trade and Industry and JAVADA determined by the Ministry of Health, Labour and Welfare. As a result of the evaluation experiment using SVM, in the evaluation item of "posture", the accuracy of language and action feature set showed 0.89, and in "assertion of opinion", the accuracy of action and head feature set showed 0.87. And, the weight of the feature quantity which contributed to the estimation was examined in order to investigate the relation between each evaluation item and multimodal information. In this paper, from these results, the relation between multimodal information and evaluation in the interview scene is described.

Scene understanding is a central problem in a field of computer vision. Depth estimation, in particular, is one of the important applications in scene understanding, robotics, and 3-D reconstruction. Estimating a dense depth map from a single image is receiving increased attention because a monocular camera is popular, small and suitable for a wide range of environments. In addition, both multi-task learning and multi-stream, which use unlabeled information, improve the monocular depth estimation efficiently. However, there are only a few networks optimized for both of them. Therefore, in this paper, we propose a monocular depth estimation task with a multi-task and multistream network architecture. Furthermore, the integrated network which we develop makes use of depth gradient information and can be applied to both supervised and unsupervised learning. In our experiments, we confirmed that our supervised learning architecture improves the accuracy of depth estimation by 0.13 m on average. Additionally, the experimental result on unsupervised learning found that it improved structure-from-motion performance.

Predicting when and where events will occur in cities, like taxi pick-ups, crimes, and vehicle collisions, is a challenging and important problem with many applications in fields such as urban planning, transportation optimization and location-based marketing. Though many point processes have been proposed to model events in a continuous spatio-temporal space, none of them allow for the consideration of the rich contextual factors that affect event occurrence, such as weather, social activities, geographical characteristics, and traffic. In this paper, we propose DMPP (Deep Mixture Point Processes), a point process model for predicting spatio-temporal events with the use of rich contextual information; a key advance is its incorporation of the heterogeneous and high-dimensional context available in image and text data. Specifically, we design the intensity of our point process model as a mixture of kernels, where the mixture weights are modeled by a deep neural network. This formulation allows us to automatically learn the complex nonlinear effects of the contextual factors on event occurrence. At the same time, this formulation makes analytical integration over the intensity, which is required for point process estimation, tractable. We use real-world data sets from different domains to demonstrate that DMPP has better predictive performance than existing methods.

The Japanese language is known to have a rich vocabulary of mimetic words, which have the property of sound symbolism; Phonemes that compose the mimetic words are strongly related to the impression of various phenomena. Especially, human gait is one of the most commonly represented phenomena by mimetic words expressing its visually dynamic state. Sound symbolism is useful for modeling the relation between gaits and mimetic words intuitively, but there has been no study on their intuitive generation. Most previous gait generation methods set specific class labels such as “elderly” but have not considered the intuitiveness of the generation model. Thus, in this paper, we propose a framework to generate gaits from a mimetic word based on sound symbolism. This framework enables us to generate gaits from one or more mimetic words. It leads to the construction of a generation model represented in a continuous feature space, which is similar to human intuition. Concretely, we train an encoder-decoder model conditioned by a “phonetic vector”, a quantitive representation of mimetic words, with an adaptive instance normalization module inspired by style transfer. The phonetic vector is a dense description of the intuitive impression of a corresponding gait and is calculated from many mimetic words in the HOYO dataset, which includes gait motion data and corresponding mimetic word annotations. Through experiments, we confirmed the effectiveness of the proposed framework.

In multi-agent environments, reinforcement learning (RL) has shown strong potential especially with the recent developments. However, there exist different difficulties in developing cooperation among learning agents in practical environment, typically that includes continuous space, partial observability and competitive situation. Therefore, in this research, we focus on the cooperative behavior on Predator-Prey game which is widely used as one of the typical simulations of Multi-agent environment. Especially, we focus on predators that their goal is to catch a prey. We propose Leader-Follower model that employs leader’s instruction and coercion, and investigate how they cooperate with each other to achieve their goal considering the prey’s policy using a model of RL. To train policies in an effective manner, we applied curriculum learning during learning process. Evaluation and analysis of cooperation in dynamic environment is also challenging. We use agents’ trajectory to evaluate predators’ team performance in addition to the total number of capturing prey. We confirmed that our proposed method can introduce effective teamwork of predators through learning.

In order to avoid overfitting when dealing with complex data with a pattern recognition model, it is necessary to remove in advance the extraordinary values that deviate significantly from the true pattern of the population. In this study, the sample space was divided into a highly versatile space and a low versatility space by using the globally optimal decision tree. Then, the space with a low evaluation value was defined as the space with relatively large noise, and the pattern recognition model was created except for the data that belongs to the space with relatively large noise.

It was found that the pattern recognition model constructed in this way can obtain the prediction accuracy higher than that of the conventional method.

In particular, when the prediction accuracy of the model was confirmed by the walk-forward method using financial time-series data under the same conditions as actual fund management, investment performance that stably exceeded the return of benchmark assets was obtained over the past 20 years.

A space containing noise or a space in which a pattern is not easily recognized is not necessarily a subset of the entire sample set divided into two on one side. Therefore, when dividing a space by a decision tree, it is desirable to subdivide the space by a multiway tree. In this study, it was confirmed that the prediction accuracy improved when the binary tree was changed to the ternary tree.

Towards close collaboration between a human and a large number of AI systems, we propose to design an AI agent with two technical elements. The first is the use of a modeling approach that enables us to know what AI agents are trying to do. The second is the use of a multi-agent consensus building algorithm. A good combination of these two, a human and a group of AI agents were put together as one team. In this paper, we explain a configuration using a Behavior Tree and a Contract Net Protocol as a concrete example. In addition, we propose a method of applying reinforcement learning in which the intentions of the AI agents can be easily grasped by a human. The effectiveness and feasibility of this approach were evaluated with teams in a simulated Tail Tag game. Matches were held with up to 29 AI agents and 1 person on one team and 30 people on the other team. The results indicate that our approach works almost evenly with human-human collaboration by sharing roles between a human and AI swarm.

An attentive listening system for autonomous android ERICA is presented. Our goal is to realize a humanlike natural attentive listener for elderly people. The proposed system generates listener responses: backchannels, repeats, elaborating questions, assessments, and generic responses. The system incorporates speech processing using a microphone array and real-time dialogue processing including continuous backchannel prediction and turn-taking prediction. In this study, we conducted a dialogue experiment with elderly people. The system was compared with a WOZ system where a human operator played the listener role behind the robot. As a result, the system showed comparable scores in basic skills of attentive listening, such as easy to talk, seriously listening, focused on the talk, and actively listening. It was also found that there is still a gap between the system and the human (WOZ) for high-level attentive listening skills such as dialogue understanding, showing interest, and empathy towards the user.

To enhance the stability of the financial markets, price limits have been implemented in numerous financial markets. The effects to markets’ stability and traders’ profitability of price limits have been discussed by previous research. But it has not been discussed when there is difference between traders on speed of information acquisition. The asymmetry of information acquisition between traders can be observed in many situations, e.g., overseas investors and domestic investors, insiders and outsiders. Because methods, languages, etc. they use to get information are different, they obtain information about the fundamental values of financial commodities with different speeds. We used a double-auction artificial market to simulate when difference on speed of information acquisition exists, how price limits effect the stability of the financial market, and the profitability of traders who has different speeds to obtain the information about fundamental value. We found if there is difference of speed for getting information between traders, price limits do not always enhance the stability of market. When the band of price limits is loose, the volatility of market price will rise if the ratio of traders with fast speed of information acquisition is relatively large. And when the band of price limits is loose, some traders who has slower speed of information acquisition may be hurt by price limits. Our work shows it is necessary to consider the difference on speed of information acquisition when designing some market institutions like price limits.

Multi-agent inverse reinforcement learning (MAIRL) is a framework for inferring expert agents’ reward functions from observed trajectories in a Markov game. MAIRL consists of two steps: the calculation of the optimal policy for reward and the update of reward based on the difference between the calculated policy and the expert trajectory. The former becomes a bottleneck because it is a multi-agent reinforcement learning (MARL) problem, which causes the non-stationary problem. Avoiding this problem, we propose the parallel coordinate descent method based MAIRL, which is an extension of maximum discounted causal entropy inverse reinforcement learning to theMarkov game. A previous method that uses coordinate descent updates one agent’s reward and policy at a time when other agents’ policies are fixed. On the other hand, the proposed method updates reward and policy for each agent in parallel and exchanges other agent policies synchronously for improving learning speed. In computer experiments, we compare the learning speeds of the previous and proposed method in the case of inferring the reward of a one equilibrium solution in two agents grid navigation. Experimental results showed that the parallelization does not always improve convergence speed, that the other agent’s policies significantly affect the learning speed, and that the learning speed is improved by parallelization when the other agent’s policies are the pseudo policy that is overwritten by the expert trajectories distribution.

Auctions have long been used to trade goods that do not have a fixed price. In recent years, with the spread of internet auctions, the number of goods that can be handled at a single auction has increased. The time required to make a bidding decision per product remains the same even when a large number of goods are traded, such as in wholesale. For this reason, a large time is required to determine the valuation in proportion to the number of goods. This causes problems for the auctioneer and the participants. It is difficult for the auctioneer to set the reserve price and for the participants to decide the amount and target of their bids. It is necessary to understand the value of a product and estimate the end price. Since there is no fixed price for a product sold in an auction, we estimate the distribution of end prices instead of point estimation. We focus on B2B luxury brand goods auctions, where the number of goods is large and the bidding decision cost per product is high. We estimate the distribution of successful bids on actual auction data of wristwatches, which have large transaction volumes in the auction. The performance of the proposed method was measured by MAE, RMSE, and MAPE, and was close to that of experts. The proposed method was able to capture the data distribution. Finally, we show that the end price distribution estimation can be used to support both auctioneers and participants in brand-name auctions with a large number of goods.

We propose a coordinated control method of agents, which are self-driving ridesharing vehicles, by using multi-agent deep reinforcement learning (MADRL) so that they individually determine where they should wait for passengers to provide better services as well as to increase their profits in rideshare services. With the increasing demand for ridesharing services, many drivers and passengers have started to participate. However, many drivers spend most of their operating time with empty vehicles, which is not only inefficient but also causes problems such as wasted energy, increased traffic congestion in urban areas, and shortages of ridesharing vehicles in less demand areas. To address this issue, distributed service area adaptation method for ride sharing (dSAAMS), in which agents learn where they should wait using MADRL, was already proposed, but we found that it does not work well under certain environments. Therefore, we propose dSAAMS* with modified input and improved reward scheme for agents to generate coordinated behaviors to adapt to various environments. Then, we evaluated the performance and characteristics of the proposed method by using a simulation environment with varying passenger generation patterns and real data in Manhattan. Our results indicate that the dSAAMS* provides better quality service than the conventional methods and performs better in dynamically changing environments.

To date, many studies concerned with intrinsic motivation in humans and artificial agents based on a reinforcement learning framework have been conducted. However, these studies have rarely explained the correspondence between intrinsic motivation and other essential cognitive functions. This study aims to build a method to express curiosity in new environments via the ACT-R (Adaptive Control of Thought-Rational) cognitive architecture. To validate the effectiveness of this proposal, we implement several models of varying complexity using the method, and we confirm that the model’s behavior is consistent with human learning. This method focuses on the“ production compilation” and ”utility” modules, which are generic functions of ACT-R. It regards pattern matching with the environment as a source of intellectual curiosity. We prepared three cognitive models of path planning representing different levels of thinking. We made them learn in multiple-breadth maze environments while manipulating the strength of intellectual curiosity, which is a type of intrinsic motivation. The results showed that intellectual curiosity in learning an environment negatively affected the model with a shallow level of thinking but was influential on the model with a deliberative level of thinking. We consider the results to be consistent with the psychological theories of intrinsic motivation. Furthermore, we implemented the model using a conventional reinforcement learning agent and compared it with the proposed method.

To design a dialogue well and easily implement it into an agent dialogue system, we constructed and verified typical dialogue patterns as “ Dialogue Design Patterns ”. To analyze the dialogues, we defined the “ Standard Dialogue Structure ”that represents simple components of dialogues. This is constructed using the“ Turn ”that is one turn-taking between the dialogue participants, the“ Topic ”that is a series of Turns in which a single question is resolved, the“Topic-shift”that is short interaction in which a consensus is built before moving to the next Topic, and the“ Scene ”that puts together each component. We investigated the frequent patterns of dialogue in the Topic. We analyzed the content of each Turn’s utterances on two axes:“convey-receive”and“logical-emotional”. These Turns were labeled on a four-quadrant impression as“ teach ”(logical convey),“ empathy ”(emotional convey),“ survey ” (logical receive), and“ listen ”(emotional receive). We analyzed the frequency of appearance for these labels in the Turns for each Topic using frequent pattern analysis, and we found several typical patterns that order these labels. We named these the patterns and determined their tendency scores. The tendencies of these patterns were categorized into“ information amount ”,“ cooperation ”, and“ balance ”. Finally, we implemented a dialogue scenario created using the Standard Dialogue Structures and Dialogue Design Patterns in our system. In the feasibility study with local government employees, the achievement rate of agent dialogue was more than 50%, and it higher than our previous system. This shows that dialogues based on Dialogue Design Patterns can facilitate agent dialogue. In conclusion, we showed that applying the Standard Dialogue Structure and Dialogue Design Patterns to dialogues in a human-agent dialogue system is a practical and possible way of easily implementing effective dialogues that facilitate interaction even for novices in dialogue design.

Recently, most stock exchanges in the U.S. employ maker-taker fees, in which an exchange pays rebates to traders placing make orders (remaining on an order book) and charges fees to traders taking orders (executed immediately). The maker-taker fees will encourage traders to place many make orders and the orders will provide liquidity to the exchange. However, the effects of the maker-taker fees for a total cost of a taking order, including all the charged fees and market impact, are not clear. In this study, we investigated the effects of the maker-taker fees for the total costs of a taking orders using our artificial market model, which is an agent-based model for financial markets. In addition, we examine the difference of market liquidity in the market between with and without a makertaker fee structure. We found that the maker-taker fees encourage the traders to provide liquidity, whereas increase the total costs of taking orders. Furthermore, we found market liquidity improved when the market maker rebates increased.

In this paper, we propose a novel method of reward design for multi-agent reinforcement learning (MARL). One of the main uses of MARL is building cooperative policies between self-interested agents. We take inspiration from the concept of mechanism design from game theory to modify how agents are rewarded in MARL algorithms. We defined the payment that reflects the negative contribution to other agents’ valuation in the same manner as the Vickrey-Clarke-Groves (VCG) mechanism. We give the individual learning agent a reward signal that consists of two elements. One is a reward evaluated solely on the basis of individual behavior that will follow a greedy and selfish policy, and the other is a negative reward as a penalty evaluated on the basis of the payment that will reflect the negative contribution to social welfare. We call this scheme reward design for MARL based on the payment mechanism (RDPM). We experimented with RDPM in two different scenarios. We show that RDPM can increase the social utility among agents while the other reward designs achieve far less, even for basic and simplistic problems. We finally analyze and discuss how RDPM affects the building of a cooperative policy.

In this paper, we investigate the acceptability of a non-task-oriented dialogue system that uses utterances to get closer psychologically. We defined utterances to get closer psychologically as “utterances that express intimacy with the other person or a favorable feeling toward the other person, such as joking, sympathy, compliment, or non-honorifics utterances (like a friend)”. Conventional research has reported that jokes, non-honorifics utterances, and compliments are useful for building a smooth relationship between a dialogue system and a user. On the other hand, individual differences in acceptability to utterances to get closer psychologically are considered to be large. In particular, we believe that the personality characteristics of the user affect the acceptability of utterances to get closer psychologically. So, we set research question 1: “How do user personality traits affect the acceptability of a non-task-oriented dialogue system with utterances get closer psychologically?” Also, utterances get closer psychologically has the risk of making the interlocutor uncomfortable. Therefore, in considering the implementation of utterances gets closer psychologically in a dialogue system, it is useful to examine how different strategies of utterances get closer psychologically affect the acceptability of a chatting dialogue system. So, we set research question 2: “How do different utterance strategies to get closer psychologically affect the acceptability of chatting dialogue systems?” To discuss these research questions, we conducted a dialogue experiment using a rule-based non-task-oriented dialogue system (n = 82). The results showed that for RQ1, among the five personality characteristics targeted in this experiment, the user’s diligence was related to the evaluation of the non-task-oriented dialogue system for utterance strategies to get closer psychologically used in this experiment in the subjective index, and extroversion, neurotic tendency, and openness in the objective index (likability based on user utterances). For research question 2, the experimental results showed that the acceptability between utterance strategies to get closer psychologically was significantly different in the viewpoint of the subjective index. These findings contribute to the design of a non-task-oriented dialogue system.

Though modern organization theory views organizational decision making from a very rational perspective, it is known that actual organizational decision-makings are often done through organized anarchy with 　“many autonomous actors operating with bounded rationality in an environment with ambiguous goals, an unclear link, between cause and effect, and fluid participation with the activities and subgroups of the organization”, which is well-described by so-called “the garbage can model.” In this study, we investigate how much the introduction of time constraints into the decision of garbage cans (opportunities) can improve the problems arised from organized anarchy. The analyses show that the introduction of time constraints can decrease the number of unsolved problems and also that the number of solved problems is maximized at some length of time constraints in specific organizational structures. These results as a whole indicate the mere introduciton of deadline may improve problems caused by organized anarchy.