Satoshi Kurihara

Adaptive Selection of Reactive/Deliberate Planning for the Dynamic Environment

This paper proposes and evaluates a methodology for multi-agent real-time reactive planning. In addition to the feature of conventional real-time reactive planning, which can react in a dynamic environment, our planning can perform deliberate planning when, for example, the robot has enough time to plan its next action. The proposed planning features three kinds of agents: a behavior agent that controls simple behavior, a planning agent that makes plans to achieve its goals, and a behavior selection agent that intermediates between behavior agents and planning agents. They coordinate a plan in an emergent way for the planning system as a whole. We confirmed the effectiveness of our planning by means of a simulation. Furthermore, we implemented an active vision system, which is the first stage of building the real-world agent, and used it to verify the real-world effectiveness of our planning.

Reusing Coordination and Negotiation Strategies in Multi-Agent Systems for Ubiquitous Network Environment

Recently, we proposed an intelligent ubiquitous computing (ubicomp) environment where sensors and/or their stations/servers have CPUs to cooperatively learn generalized series of sensed events that are involved in human activities. This can be regarded as a multi-agent application. Because ubicomp applications target support for daily-life activities, one of their characteristics is that the same/similar series of events occurs frequently. Multi-agent plans in applications of this type are used to foresee human activities and generate programs to assist them. Therefore, the same planning processes for conflict detection and resolution recur. This paper proposes a learning method in which past plans are exploited for problem solving in an environment where the same/similar problems appear repeatedly. We discuss how the plan is stored and reused using as an example the exploration of conflict-free routes in a room and then describe experimental results.

On the use of hierarchical power-law network topology for server selection and allocation in multi-agent systems

In this paper, we focus on the effectiveness of using the power-law relationship that appears in actual network topology for solving server selection and allocation problems in multi-agent systems (MAS). We introduce the reverse weighted degree (RWD) server selection algorithm, which selects the nearest server with a lower load average, and the concept of the scope, which spreads the range of the topological information about neighbors. Furthermore, we evaluate the efficiency and fairness of the algorithm when server deployment is performed by using the degree-oriented server allocation, which places the server agent on more convenient nodes from the viewpoint of network topology, and by random server allocation. From simulation results using the real Internet topology, we find that awareness of the network structure can improve the total performance of agents significantly, though previous approaches in MAS did not consider the topology of the network.

Effective Planning for Conflicting Situations for Ubiquitous Sensor Network Environments

Applications of sensor networks and ubiquitous computing have received attention. They can provide many kinds of important services for supporting daily and social activities in home, schools, offices and public spaces in the future (Kurihara, 2008). However, to realize these kinds of applications, a number of new technologies in AI and multi-agent systems (MAS) are also required because many devices and control programs are concurrently work to achieve their goals in cooperation with other ones. These works arise according to the human requirements based on their individual activities. In order to achieve these required goals, each agent has to create the plan (means-end analysis) and then performs it. However, the plan often conflict with those that are being created, already being scheduled, and executed by other agents because of the limited resources. Furthermore, since the human’s activities are usually real-time with deadline, the agent must also be able to complete its planning and resolution of these conflicts within a reasonable time to have an acceptable quality plan. This means that both efficient planning and sophisticated conflict resolution are strongly required. We adopt hierarchical planning (for example, see (Erol & Nau, 1994; Sacerdoti, 1974) using the decision-theoretic planning approach (Goldwin, & Simmons, 1998) for efficient planning but it is not trivial to apply hierarchical planning to MAS. In hierarchical planning, appropriate (abstract) plans are selected level by level to maximize the utility U(p), where p is the expected final plan comprising a sequence of primitive actions. However, in the MAS context, conflicts between agents affect the efficiency and quality of resulting plans. When a conflict is found at lower levels, an additional sophisticated process for avoiding it (conflict resolution) must be invoked and some extra actions (such as waiting for synchronization and detouring) may have to be added to the plan. The conflict resolution process may become costly or fail. Even a single conflict, if it is difficult to resolve, will result in a plan with 5

How Collective Intelligence Emerge in Complex Environment

In this paper we analyze a simple adaptive model of competition called the Minority Game, which is used in analyzing competitive phenomena in markets. The Minority Game consists of many simple autonomous agents, and self-organization occurs as a result of simple behavior rules. Up to now, the dynamics of this game have been studied from various angles, but so far the focus has been on the macroscopic behavior of all the agents as a whole. We are interested in the mechanisms involved in collaborative behavior among multiple agents, so we focused our attention on the behavior of individual agents. In this paper, we suggest that the core elements responsible for forming self-organization are: (i) the rules place a good constraint each agent’s behavior, and (ii) there is a rule that leads to indirect coordination. Moreover, we tried to solve the El Farol’s bar problem based our suggestions.

Simple but efficient collaboration in a complex competitive situation

In this paper, we analyze a simple adaptive model of competition called the Minority Game, which is used in analyzing competitive phenomena in markets, and suggest that the core elements required for the formation of self-organization are: (i) rules that place a good constraint on each agents behavior, and (ii) rules that lead to indirect coordination, which is called stigmergy. Finally, we tested the points suggested by this research in solving the El Farols bar problem, which is an extended version of the Minority Game.

Characteristics of Contract-Based Task Allocation by a Large Number of Self-Interested Agents

Task and resource allocation is a key technology in Internet applications, such as grid computing and agent grids, for appropriately allocating tasks such as video and music downloads, scientific calculations, and language services (Cao et al., 2005; Chunlin & Layuan, 2006). In these applications, tasks should be allocated to and done in appropriate agents. Thus negotiation protocol for task allocation is one of the important issues in multiagent systems research. In particular, the contract net protocol (CNP) and its extensions have been widely used in certain applications because of their simplicity and superior performance (Sandholm, 1993; Smith, 1980; Weyns et al., 2006). Agents in CNP play one of two roles: as managers that are responsible for allocating tasks and monitoring processes, or as contractors that are responsible for executing the allocated tasks. A manager agent announces a task to contractor agents, which bid for the task with certain promised values (such as cost, duration, and payment). The manager then awards the task to the contractor (called an awardee) that bid with the best tender (before the deadline) and allocates the task to it. However, the naive CNP is usually assumed to be applied to allocating tasks in smallscale, non-busy multi-agent systems (MASs). Interference among agents is always observed in this kind of negotiation protocol. Consider many agents having to allocate tasks to other efficient contractors. In basic CNP, a contractor agent that receives task announcements bids for the tasks one by one. When many tasks are announced by many managers, however, they have to wait a long time to receive a sufficient number of bids. In the original conception of CNP (Smith, 1980), multiple bids were proposed to concurrently handle many announcements. If a contractor is awarded multiple tasks simultaneously, however, it may not be able to provide its promised quality or performance. In fact, efficient contractor agents are selected as awardees by many manager agents, leading to a concentration of tasks. In addition, when a large number of O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg

An Implementation for Capturing Clickable Moving Objects

This paper discusses a method for identifying clickable objects/regions in still and moving images when they are being captured. A number of methods and languages have recently been proposed for adding point-and-click interactivity to objects in moving pictures as well as still images. When these pictures are displayed in Internet environments or broadcast on digital TV channels, users can follow links specified by URLs (e.g., for buying items online or getting detailed information about a particular item) by clicking on these objects. However, it is not easy to specify clickable areas of objects in a video because their position is liable to change from one frame to the next. To cope with this problem, our method allows content creators to capture moving (and still) images with information related to objects that appear in these images including the coordinates of the clickable areas of these objects in the captured images. This is achieved by capturing the images at various infrared wavelengths simultaneously. This is also applicable to multi-target motion capture.

Learning implicit resource relationships from past plans in multi-agent systems

This paper discusses storing and analyzing the hierarchical planning results in the past in order to identify implicit costs and resource relationships between activities in multi-agent contexts.