Effective agent coordination is crucial in cooperative Multiagent Reinforcement Learning (MARL). While recent advances have significantly improved cooperation by modeling agent interactions through various graph structures, most existing approaches primarily focus on homogeneous agents. Despite the ubiquity of heterogeneous agents, constructing a comprehensive graph that captures their diverse attributes and relationships from scratch is notoriously labor-intensive for both humans and agents, which makes policy learning extremely challenging. To tackle this difficulty, we propose a novel method that utilizes a fuzzy human attention-guided graph to model inter-agent relationships. Instead of learning the graph entirely from scratch, we incorporate abstract human attention, with its uncertainty captured through fuzzy logic, to guide the graph development process. To further accommodate the varying attributes and objectives of heterogeneous agents while maintaining their learning capabilities, the attention-guided graph is fine-tuned through a hyper-network. Our proposed approach is end-to-end trainable and agnostic to specific MARL methods. Empirical evaluations conducted on challenging heterogeneous scenarios from the StarCraft Multiagent Challenge (SMAC) and SMACv2 validate the effectiveness of the proposed method.

Abstract Due to its exceptional learning ability, multi-agent deep reinforcement learning (MADRL) has garnered widespread research interest. However, since the learning is data-driven and involves sampling from millions of steps, training a large number of agents is inherently challenging and inefficient. Inspired by the human learning process, we aim to transfer knowledge from humans to avoid starting from scratch. Given the growing emphasis on the Human-on-the-Loop concept, this study focuses on addressing the challenges of large-population learning by incorporating suboptimal human knowledge into the cooperative multi-agent environment. To leverage human experience, we integrate human knowledge into the training process of MADRL, representing it in natural language rather than specific action-state pairs. Compared to previous works, we further consider the attributes of transferred knowledge to assess its impact on algorithm scalability. Additionally, we examine several features of knowledge mapping to effectively convert human knowledge to the action space where agent learning occurs. In reaction to the disparity in knowledge construction between humans and agents, our approach allows agents to decide freely which portions of the state space to leverage human knowledge. From the challenging domains of the StarCraft Multi-agent Challenge, our method successfully alleviates the scalability issue in MADRL. Furthermore, we find that, despite individual-type knowledge significantly accelerating the training process, cooperative-type knowledge is more desirable for addressing a large agent population. We hope this study provides valuable insights into applying and mapping human knowledge, ultimately enhancing the interpretability of agent behavior.

Text-based Computer-Mediated Communication (CMC) diminishes the social presence of the conversational partner owing to the lack of nonverbal cues exchanged in face-to-face (FtF) communication. This study aims to enhance social presence in a dyadic text chat by employing a robot avatar that supplements nonverbal cues. In particular, we propose a chat system with a robot avatar that can express gestures in response to the user’s actions (typing, sending messages) and read the user’s messages aloud. In the evaluation experiment, nine pairs of participants (for a total of 18 participants) used two types of chat system: the proposed system and the baseline system in which a robot avatar only reads messages aloud. As a result, the proposed robot avatar significantly increased social presence compared to the baseline system. Furthermore, the proposed gesture expressions significantly improved the ease of chatting. Specifically, our results suggest that robot gestures in response to typing might have an effect similar to nonverbal cues in FtF communication.

With the development of large-scale language models, research on human-AI interaction is gaining momentum. Our aim is to encourage users to disclose information that can lead to health benefits through interaction with AI. To achieve this, we focused on using various facial expressions with emojis, and explored methods to reduce discomfort when speech is expressed simultaneously with these facial expressions. As a result, although an upward trend in positive emotions was observed, it became clear that more detailed analysis of the text used in conjunction with emojis is necessary for further improvement.

Most research investigating how social influence affects election results mainly uses diffusion models for binary opinions. However, these diffusion models are progressive and focus on the diffusion of one opinion. In this article, we introduce a general diffusion model for ordinal opinions expressed as linear orderings over a finite set of candidates. We employ agent-based modeling to simulate a nonprogressive diffusion process, allowing multiple types of opinion diffusion about different candidates. The proposed agent-based diffusion model can forecast long-term trends of opinion diffusion in social networks by capturing voters’ personalized features and incorporating dynamic social contexts. Furthermore, we examine the possibility of affecting election outcomes by externally changing the ordinal opinions of certain vertices, i.e., campaigning. Since finding influential voters from the social network is computationally challenging, we propose a heuristic approach, i.e., backward influence rank (BIR). Experimental results demonstrate that the proposed BIR approach is superior to the classic greedy approach for campaigning by achieving a similar margin of victory to that of the greedy approach but running two orders of magnitude faster than the greedy approach did.

This research explores the application of deep learning techniques, specifically employing a residual neural network, to predict Kellgren-Lawrence grade (KLG) in osteoarthritis patients using magnetic resonance images (MRI). Taking advantage of the characteristics of images, the proposed model integrates the MRI slice number and the use of segmented images. Unlike conventional approaches, we adopt a one-to-one image processing strategy, so our model takes each slice individually as input and returns a prediction for each of them to enhance the model’s ability to focus on specific slices and increase the results’ interpretability. Furthermore, results on real-world data corroborate the idea that the segmented image can provide more accurate prediction by allowing our network to focus on the crucial parts of the knee. The empirical results show the model’s promising performance in predicting KLG, demonstrating its potential for accurate and detailed diagnosis of osteoarthritis. This research contributes to advancing studies on the early prediction of osteoarthritis by proposing an effective and interpretable deep-learning framework for osteoarthritis assessment.

Early detection and treatment of diseases are crucial in preventing and controlling their progression. Developing methods for a super-early screening is highly desirable. We have continuously conducted health survey sessions for elderly adults, and have collected information on their health, such as their living environment and physical functions. This paper presents the design of the measurement session. At the latest measurement session, we focus on DLB (Dementia with Lewy Bodies). We then try to construct a screening method for Parkinsonism, which is one of the prodromes of DLB. In the survey, we collected not only health information but also gait measurements. It attempts to realize a highly accurate screening tool with modality fusion; each modality is based on the gait analysis, in addition to measurements, questionnaires, and interviews. Next, we show the difference between the active and inactive groups in the last five sessions of information. The results show each of the two groups made contact with different types of communities. The results also show that the active group excelled in activities of daily living and working memory and had few concerns about dementia.

Due to the exponential growth of agent interactions and the curse of dimensionality, learning efficient coordination from scratch is inherently challenging in large-scale multi-agent systems. While agents' learning is data-driven, sampling from millions of steps, human learning processes are quite different. Inspired by the concept of Human-on-the-Loop and the daily human hierarchical control, we propose a novel knowledge-guided multi-agent reinforcement learning framework (hhk-MARL), which combines human abstract knowledge with hierarchical reinforcement learning to address the learning difficulties among a large number of agents. In this work, fuzzy logic is applied to represent human suboptimal knowledge, and agents are allowed to freely decide how to leverage the proposed prior knowledge. Additionally, a graph-based group controller is built to enhance agent coordination. The proposed framework is end-to-end and compatible with various existing algorithms. We conduct experiments in challenging domains of the StarCraft Multi-agent Challenge combined with three famous algorithms: IQL, QMIX, and Qatten. The results show that our approach can greatly accelerate the training process and improve the final performance, even based on low-performance human prior knowledge.

In multi-agent systems, agents with limited capabilities need to find a cooperation partner to accomplish complex tasks. Evaluating the trustworthiness of potential partners is vital in partner selection. Current approaches are mainly averaged-based, aggregating advisors’ information on partners. These methods have limitations, such as vulnerability to unfair rating attacks, and may be locally convergent that cannot always select the best partner. Therefore, we propose a ranking-based partner selection (RPS) mechanism, which clusters advisors into groups according to their ranking of trustees and gives recommendations based on groups. Besides, RPS is an online-learning method that can adjust model parameters based on feedback and evaluate the stability of advisors’ ranking behaviours. Experiments demonstrate that RPS performs better than state-of-the-art models in dealing with unfair rating attacks, especially when dishonest advisors are the majority.

We have proposed a novel sampling method (NS) for controlling step sizes and incorporating Lévy flight to distribution-based discrete particle swarm optimizations (DDP-SOs), which are discrete extended variants of particle swarm optimizations handling the continuous parameters of probability distributions over the variable values instead of directly handling discrete variables. Our previous work demonstrated that NS improved all DDPSOs on function optimization problems. However, on categorical problems, the NS did not improve DDPSOs designed for integer problems. For more detailed investigations, we conducted optimization experiments on NK landscapes. The results show that NS improves the DDPSOs' efficiency and robustness to large solution space and non-separable problems. Besides, we found that NS is effective on integer DDPSOs even for categorical optimization in cases where decision variables have a few states. In addition, the proposed methods were tested on feature selection experiments and achieved superior results compared to some evolutionary algorithms designed for feature selection.

In recent years, Social Networking Services (SNS) have become popular among young people. Unfortunately, as SNS usage has increased, cyberbullying has also increased and has become a social problem. Several previous studies have employed multi-agent simulation, which can be used to analyze human relationships, to identify bullying mechanisms. In this study, we model SNSs in a classroom and apply multi-agent simulation to analyze the influence of SNSs on classroom friendships. We focus on junior high school students as our research object. In the proposed model, which is based on socion theory, an agent can communicate with other agents using two types of networks: a classroom network and SNS networks, via a network which recognized by each individual (in socion theory, people have mental networks that reflect society). Agents communicate face to face (FTF) in the classroom and using SNS in their SNS groups. In addition, agents have social skills and are categorized based on these social skills. In this study, we simulate friendship relations considering SNSs and discuss the influence of SNSs on classroom relationships based on the simulation results. We performed two simulations; one only involved FTF communication and the others involved both FTF and SNS communication. We compare the two simulations and discuss the results. We found that, compared to FTF communication, the average likability rating of agents increased with SNS communication. On the other hand, we also found that specific agents were rejected. We consider that sharing information over SNSs is related to increased bullying. In conclusion, we discuss applying to educational robots from results.

Nonindependent mate choice occurs when a female is in-fluenced in her choices of mate by the social environment. Frequency-dependent selection (FDS) is a typical example of a nonindependent mate choice and comes in two forms: positive or negative. In the positive form, any rare variant is at a disadvantage, whereas rare variants are favored in the negative form. Both forms of FDS have been confirmed in many species, and several mathematical and theoretical biol-ogy studies have reported the advantages of each. However, few studies have focused on the evolution of the two forms of FDS together. In this work, we simulated FDS using an agent-based model consisting of imported mating strategy as gene and female preference influenced by the social environ-ment as meme. Experimental results revealed a relation be-tween the operational sex ratio of males and the FDS strategy of females. A similar tendency was observed among real an-imals.

This paper presents a multiagent-based simulation approach to the effect of network structure difference on the diffusion of items. Recently, the rapid spread of information and communication technology induces multiplexing of our communication space. As a result, diffusion of products in the market has been changed because the communication affects our behaviors and state of own mind. Network externality is an effect that value of a product depends on the market penetration. It has been known that markets of products having network externality are greatly influenced by interpersonal communication. In this paper, we constructed two network, “offline network” and “online network” that refer to networks before and after the development of information and communication technologies, focusing on such changes and analyzed the differences between network structures. We verified the effect that difference of network structures affects the diffusion of items in network effect markets. We discussed the property of diffusion process in each network and how easily monopolistic diffusion can occur depending on the network structures. Introduction Recently, a socializing method and other parties of association have been changed because of the development of information and communication technologies such as the Internet and mobile phones. For example, mobile phones make it possible to communicate with anyone, anytime, and the Internet and Social Network Service (SNS) make it easy for people who have something in common to communicate through online communities. It is thought that communication and exchanging information with others, and their behavior, affects our own actions and psychology. The network effect (also called network externality or demand-side economies of scale) in markets is an example of interaction with others producing a major effect on individuals (Katz and Shapiro, 1985). Briefly, the network effect means that as the number of users an item has increases, that item becomes more attractive to others. In this situation, it is thought that the diffusion rate of items within one group of friends affects the decision about whether or not to buy something. When two similar items are in a race to be the most popular, it is common for only one item to be shared exclusively because of positive feedback and items selling faster if they are already doing well. This phenomenon is called “winnertakes-all” by Arthur (1996). There has been much previous research about the network effect (e.g., Ozawa and Nakayama (2010); Weitzel et al. (2003)). Kaneko et al. (2006, 2005) constructed a model that considered the asymmetricity of information and analyzed customers’ purchase decision-making processes. In the model each consumer had different information with respect to others’ purchase behavior, and they reported that the market becomes inefficient if customers are unaware of each other’s behavior. Kawamura and Ohuchi (2005) analyzed the effectiveness of the present strategy, in which businesses provide their services without charge. They examined the effectiveness of two present strategies: a simple present strategy and friend present strategy, and discussed the effectiveness of present strategies in different network structures. Iba et al. (2001) analyzed the format competition of video cassette recorders, that is to say a typical standard race interaction between customers, by the artificial market model with multiagent approach. The simulation observed the emergence of locality, which is caused by the local influence, and the results showed that the local clusters provide the brakes on the winner-take-all phenomenon. Mizutani (2002) proposed a simulation model with evincive individual relationships. He used network structures that describe the relationship in urban and provincial areas, and showed that differences to these structures can affect the diffusion race. In a network effect market, the situation at the early stage is significant in terms of the final diffusion result (Liebowits and Margolis, 1994). Therefore, we need to examine the early stage and come up with a definition for the customer group that decides to buy an item at the early stage and how the diffusion process proceeds from that point. Additionally, it seems to be important to study acquaintance network structures that show the relationships between customers. We propose a model that draws on Mizutani (2002)’s model that focuses on the relationship between network structure differences and the diffusion state. First, we deALIFE 14: Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living Systems

This paper presents a verbally-based cognitive task for elderly with mild cognitive impairment. As designed with conscious of daily conversation, the task is done by oral answering some questionnaire. An elderly firstly talks about the topics of favorite season, travel, gourmet, and daily life, and then he/she does three cognitive tasks of reminiscence task, category recall, and working memory task. With the use of the functional near-infrared spectroscopy (fNIRS), which can measure cerebral blood flow activation non-invasively, we had collected 42 CHs fNIRS signals on frontal and right and left temporal areas from 22 elderly participants (7 males and 15 females between ages of 64 to 89) during cognitive tests in a specialized medical institute. All participates are classified into three clinical groups: elderly individuals with cognitively normal controls (CN), patients with mild cognitive impairment (MCI), and mild Alzheimer's disease (AD). Toward personalized cognitive training, we report a task effect measurement by the statistical tests of fNIRS signals.

Differential Evolution (DE) is a powerful stochastic algorithm for real-coded optimization. However, DE has a problem as well as other traditional stochastic optimization algorithms: difficult to optimize search spaces that are little globally convex. Thus, it is difficult for DE and traditional algorithms to optimize some practical problems where globally convex cannot be assumed. As one of the solution for this problem, we propose Differential Evolution on Scattered Parents (DE-SP) that re-selects the individuals on each dimension when the mutant individual is calculated and some children individuals' candidates unconditionally become the children individuals. We have implemented 2 types of optimization experiment to verify the performance of DE-SP: Noisy Function 1 (NF1), that is a benchmark problem with little globally convex, Noisy Function 2 (NF2), that is the one with globally convex. And we compared the performance of DE-SP with those of 15 types of algorithms. Thereby, it is confirmed that DE-SP was the most stable algorithm to optimize little globally convex spaces among 15 comparative algorithms from the experiment of optimizing NF1. And it is comfirmed that DE-SP can optimize globally convex spaces as well as DE from the experiment of optimizing NF2.

A method for adding a target emotion to arbitrary body movements of a human form robot (HFR) is developed. The additional emotion is pleasure, anger, sadness or relaxation. This paper proposes a motion rendering system that modifies arbitrary basic movements of a certain real HFR to add the target emotion at intended strength. The system is developed on the assumption that movements can be emotive by processed on the basis of the correlations between movement features and expressed emotions. The movement features based on Laban movement analysis (LMA) are adopted. An experiment using a real HFR are conducted to test how well our system adds a target emotion to arbitrary movements at intended strength. The results of experiments suggest that our method succeeded in adding a target emotion to arbitrary movements.

The paper describes a comparison of the sensibility of recognizing emotions from human voices speaking Japanese and Korean. Our study focuses on the emotional elements included in the human voice, and our method uses Bayesian networks of prosodic features as models of Japanese's and Korean's sensibilities in recognizing emotions. The training datasets are prosodic features extracted from emotionally expressive voice data in the two languages. Our method makes the Bayesian network learn the dependence and its strength between nonverbal voice features and its emotion. We compare the sensibilities of emotion recognition from Japanese and Koreans speech by examining the cross-inference through two Bayesian networks with speech in the other language.

A central pattern generator (CPG) and passive dynamic walking (PDW) have attracted much attention in the research field of bipedal locomotion. We describe a motion control method based on dynamic joint passivization for biped robot locomotion. CPG-based motion control is effective for walking on uneven terrain. However, it has serious problems with energy loss. In contrast, PDW saves energy because a robot can walk without any active control or energy input on a downhill slope. However, PDW robot can not walk on uneven terrian, but only on a downhill slope. We think that active walking needs to be mixed with PDW for robot walking. Our motion control method is based on a mixture of the CPG and PDW, that is, the dynamic passivization of joint control. Experiments using the motion control method based on dynamic passivization of joint control successfully generated energy efficient walking and enabled superior gaits.

Recently, as the relationship between robot and human has become closer, humans demand that robots pose familiar human-like characteristics. For a robot to live and communicate with people, it requires its own personality or individuality. Changing the mood transition of robots can change the perceptions people have of their characteristics. We propose an emotion generation model that represents a robot's internal state. This model can assess the robot's individuality through mood transitions. We report experiments of emotional conversation with a robot that had this model installed. The experimental results showed that personality could be effectively expressed by changing robot's mood transitions. We also report significant results of evaluations of psychological impact.

To improve face-to-face interaction with robots, we developed a model for generating interactive facial expressions by using a simple recurrent network (SRN). Conventional models for robot facial expression use predefined expressions, so only a limited number of expressions can be presented. This means that the expression may not match the interaction and that the person may find the expressions monotonous. Both problems can be overcome by generating expressions dynamically. We tested this model by incorporating it into a robot and comparing the expressions generated with those of a conventional model. The results demonstrated that using our model increases the diversity of face-to-face interaction with robots.

A knowledge-based approach to imitation learning of motion generation for humanoid robots and an imitative motion generation system based on motion knowledge learning and reuse are described. The system has three parts: recognizing, learning, and modifying parts. The first part recognizes an instructed motion distinguishing it from the motion knowledge database by the hidden Markov model. When the motion is recognized as being unfamiliar, the second part learns it using dynamical movement primitives and acquires a knowledge of the motion. When a robot recognizes the instructed motion as familiar or judges that its acquired knowledge is applicable to the motion generation, the third part imitates the instructed motion by modifying a learned motion. This paper reports some performance results: the motion imitation of several radio gymnastics motions.

Recently, a multiuser MIMO system has attracted much attention. In the downlink of the multiuser MIMO system, since the base station simultaneously transmits signals to terminals, there is inter-user interference (IUI) at each terminal. Block diagonalization, which can achieve perfect IUI suppression, has been extensively studied to solve the issue. However, with the scheme we cannot obtain extra diversity gain due to the null steering at the base station. In this paper, we propose imperfect block diagonalization based on maximum eigenvectors of users and Gram-Schmidt orthonormalization, and its error rate performance is evaluated using computer simulations. The result shows that, despite its low complexity, the proposed scheme provides excellent performance, especially in many-user environments.

We describe a method for sensitivity communication robots which infer their dialogist's emotion. The method is based on the Bayesian approach: by using a Bayesian modeling for prosodic features. In this research, we focus the elements of emotion included in dialogist's voice. Thus, as training datasets for learning Bayesian networks, we extract prosodic feature quantities from emotionally expressive voice data. Our method learns the dependence and its strength between dialogist's utterance and his emotion, by building Bayesian networks. Bayesian information criterion, one of the information theoretical model selection method, is used in the building Bayesian networks. The paper finally proposes a reasoner to infer dialogist's emotion by using a Bayesian network for prosodic features of the dialogist's voice. The paper also reports some empirical reasoning performance.

This paper presents a method that enable a domestic robot to show emotions with its facial expressions. The previous methods using built-in facial expressions were able to show only scanty face. To express faces showing various emotion, (e.g. mixed emotions and different strengths of emotions) more facial expressions are needed. We have therefore developed a system that converts emotions into robot's facial expressions automatically. They are created from emotion parameters, which represent its emotions. We show that the system can generate facial expressions reasonably

In this paper, we propose a new behavioral model based on the concept of "meme" and "qualia" for multi-agent system. We, then, implement the model, and perform multi-agent simulations of cultural transmission. In the simulations, we construct an artificial society, where agents behave for food objects collectively. An agent transmits meme, the object to eat and its preference as a culture, to other agents. As the results of the simulations, we have observed some cultural transmissions in the artificial society.

Real-coded genetic algorithms (GAs) are effective methods for function optimization. Generally speaking, the major crossover methods used in real-coded GAs require a large execution time for calculating the fitness of many children at each crossover. Thus, a new crossover method is needed for searching such a large search space efficiently. A novel crossover method that generates children stepwise is proposed and applied to the conventional generation-alternation model. In experiments based on standard test functions and actual problems, the proposed model found an optimal solution 30-50% faster than did the conventional model.

A novel robot, Ifbot, which can communicate with humans by joyful conversation and emotional facial expression has been developed by our industry-university joint research project. This work introduces Ifbot and its mechanics and software architecture for the human-robot communication. As the fundamental technology of Ifbot for the robot-human interaction, we also propose an image processing method for real-time face tracking and talker distinction.

Many existing methods of reinforcement learning have treated tasks in a discrete low dimensional state space. However, the smooth control of humanoid robots requires a continuous high-dimensional state space. In this paper, to treat the state space, we proposed an adaptive allocation method of basis functions for reinforcement learning. Grid or incremental allocation methods have previously been proposed for allocation of basis functions. However, these methods may result in the curse of dimensionality, and a fall into local minima. On the other hand, our method avoids local minima, which are assessed by the trace of activity of basis functions. That is, if the current state is determined to have fallen into a local minimum, our method eliminates a basis function, which most affects the state. Moreover our method learns with a low number of basis functions because of the elimination process. In order to confirm the effectiveness of our method, by using computer simulation, a humanoid robot learned the motion of standing up from a chair. This motion was enabled with a small number of basis functions.

The "Ifbot" robot communicates with people by considering its own "emotions" and theirs. Ifbot has many facial expressions to communicate enjoyment. We first attempted to extract characteristics of Ifbot's facial expressions by mapping these to its emotional space, which we discuss in this paper. We applied a five-layer perceptron to the extraction. We also propose a method of seamlessly changing facial expressions using the emotional space. We report on some results of facial changes obtained with the proposed method.

Genetic algorithm (GA) is an effective method of solving combinatorial optimization problems. Generally speaking most of search algorithms require a large execution time in order to calculate some evaluation value, especially in real-world applications as well. Crossover is very important in GA because discovering a good solution efficiently requires that the good characteristics of the parent individuals be recombined. The multiple crossover per couple (MCPC) is a method that permits a number of children for each mating pair, and MCPC generates a huge amount of execution time to find a good solution. This paper proposes a novel approach to reduce time needed for fitness evaluation by "prenatal diagnosis" using fitness prediction. In the experiments based on actual problems, the proposed method found an optimum solution about 50% faster than the conventional method did. The experimental results from standard test functions show that the proposed method is applicable to other problem as well.

Recently, market and user preferences have been drastically changing; user oriented production and logistic systems such as the supply chain management system have been required. In this kind of system, different subsystems with heterogeneous level of reliability have to coexist while trying to preserve the one click response and the flexibility in order to satisfy the frequently changing and heterogeneous preferences of the users. Thus assurance is required. An autonomous decentralized database system is proposed in order to realize the assurance. In this architecture, every site can autonomously control its own database within a permissible value, and coordinate among the sites by a mobile agent. Moreover, the autonomous consistency coordination technique is proposed in order to solve the problem of integrating and keeping the consistency of DB systems with heterogeneous levels of reliability. This technique is based on the inherent properties of autonomous decentralized systems (ADS) to attain the on-line property by the use of a data field among the replicated sites. Furthermore, an evaluation of the proposed technique is showed.

Due to the advancement in the Information Technology, different kind of companies with heterogeneous needs have had the necessity to cooperate among them to get more benefit in a continuous changing market. That is heterogeneous e-business. In order to cope with such kind of e-business, a platform which can effectively realize heterogeneous needs, real time, flexibility and fault-tolerance is needed. This paper focuses on the database field that is indispensable for the e-business. Thus, it proposes Autonomous Decentralized Database System, which is composed of two techniques. In the first one, the concept of Allowable Volume (AV) is introduced, while in the second one, a Mobile Agent (MA) that adjusts the AV of each site is shown. Since the MA is always moving among the sites, and each site can always negotiate AV with the MA autonomously, the system can realize assurance. Furthermore, the effectiveness of the proposed system is shown by simulation.

By integrating heterogeneous systems with different individual requirements, it becomes possible for application systems to use the various data generated in the heterogeneous systems. Integrated systems also make it possible to process the applications from different points of view and to obtain higher assurance levels. The important requirements for the integration of application systems are not to halt the entire systems, not to violate individual system missions and to derive a synergetic effect. The autonomous decentralized system (ADS), by its online property, enables the integration of systems without stopping any system and without violating any system's mission. To collect consistent data combinations from different systems each subsystem needs to determine the combination independently and autonomously. This paper proposes a new concept, the data alliance, to determine consistent data combination from heterogeneous systems. In addition, the fairness must be maintained between the subsystems to achieve the consistency to satisfy the nature of the autonomous coordinability. This paper defines the fairness among systems in terms of the cost of the data freshness and the alliance waiting time. The data freshness value expresses the degree of how recent the data combination is for the subsystem, and the alliance waiting time expresses how long the subsystem must wait to get agreement among cooperating applications. The last part of the paper shows by simulation that the proposed technique is effective in different system topologies.