<back to Home>
<Research Plan> with Regard to the Application for the Position of Assistant Professor
System Development Tool to Create Visual Environments for Simulating and Training Collaborative Multi-Agent Systems
Multi-Agent Systems research is considered as a subfield of AI that aims to develop software agents with the ability to interact and communicate as a team to achieve more than they could individually . Multi-agent systems have recently been widely studied and implemented in business models ,  and robotic sciences , , , since developing a system with multiple agents can be highly beneficial owning to the high-speed parallel computation and low-cost scalability. Although this research topic has become increasingly significant in the highly automated world, few workable tools are available to help researchers to analyze and supervise the multi-agent teamwork. Therefore, I proposed a system development tool to create simulation environments for visualizing multi-agent systems with humaninvolved training. In developing this tool, I focus on visible agents in simulating robotic systems, but the fundamentl idea of rulebase manipulation can be extended to other business agent systems.
Fig. 1. Modules inside and outside Software Agents: (a) A general model of the software agent includes databases of beliefs and goals and a rulebase of plans. The conceptual elements of action and percept form the interface between the agent and its environment; (b) The RoboCup simulation package provides a Mediate Server (the environment), a Monitor, and a set of server-communication APIs for users to develop intelligent agents.
A. Relative Prior Work
Figure 1(a) shows a general model of the software agent. The conceptual elements of action and percept form the interface between the agent and its environment. Databases of beliefs and goals for the agent are designed along with a rulebase of plans. These modules support decision-making when events are recognized. An agent must generate an action in response to an event. The agent is typically situated in a rapidly changing environment where it has a limited view of overall situations. Therefore, directly reacting to events is also permitted before a thorough planning can be made. The practical issues of agents are diverse and may support different aspects of work; nevertheless, the multi-agent systems can be categorized into four classes based on their inter-communication conditions:
1) BDI (Belief, Desire, Intention) Model :
BDI is one of the most popular agent architectures. It has its roots in philosophy, and implements agents logically by clearly declaring their beliefs, desires, and intentions. The BDI model builds a collection of individual expert systems, rather than a team for collaboration. Although the BDI model has no coordination mechanism among the agents, it can realize swarm-type cooperation in case the agents are responsible for different parts of a series of work or simply share the load of work.
2) ALLIANCE Model :
Like the BDI model, the ALLIANCE model does not implement underlying communication among the agents. Furthermore, it has neither centralized store of world knowledge nor central executive authority. The collaboration among the agents depends on a behavior-based framework. Each agent can detect all the activities of others, and place them into the database as beliefs. Since the teamwork is achieved on the basis of ¡§implicit collaboration¡¨, the basic assumption is that the agent does not cheat the others by performing any action against its desires.
3) STEAM (Shell for TEAMwork) Model :
The STEAM model uses explicit communication to ensure the quality of coherence. Agents in this model explicitly represent their commitments and plans for joint goals based on the joint intentions theory . The development of an agent team becomes feasible, but the heavy communication leads to unavoidable overhead making the system expensive and inflexible. Therefore, the STEAM model further supports communication selectivity and team reorganization capabilities.
4) PTS (Periodic Team Synchronization) Model : The PTS model integrates the basic concepts of the above two models. Like agents in the ALLIANCE model, PTS agents collaborate with each other by following pre-determined protocols, called ¡§locker-room agreements¡¨, which are remembered identically by all agents and make their implicit collaboration efficient. Periodically, the team synchronizes with no restrictions on communication, i.e., the agents can inform each other of their entire internal states and decision-making mechanisms. Failures in belief and goal databases can thus be detected and recovered. The locker-room agreements can also be exchanged during periods of full communication if necessary. Therefore, the agent version can be upgraded quickly.
As stated above, the field of agents is large and diverse. Agent systems continuously emerge in new communication frameworks, infrastructure (e.g. brokers, yellow pages), mobility and planning approaches. Development of multi-agent systems is currently a flourishing research topic, but few integrated environments are yet available for developing, simulating and analyzing multi-agent teamwork. ¡§Virtual Environment¡¨ is an associated research area . Certain applications, mainly in visual simulation (driving simulators  and flight simulators ) and scientific visualization , , are in daily use, and are judged to be sufficiently effective to warrant further attention. These applications support human interaction well, but cannot further fit the requirement in MAS research without extending them to support the following functions: (a) generating frame codes for feasible development of individual agents; (b) uploading individual agents into the environment, and (c) recording and analyzing the multiagent teamwork. These three design issues are listed with reference to the functionality provided by the simulation league of the ¡§RoboCup¡¨ (Robot World Cup) project , . In the simulation league, RoboCup suggests a simulation package that is well-known for simulating multi-agent teamwork in playing soccer games. As demonstrated in Fig 1(b), the RoboCup package provides a Mediate Server (the environment), a Monitor, and a set of server-communication APIs for users to develop intelligent agents (software players). Users can designate the IP/Port of a Mediate Server for uploading the agents, and then the agents communicate with the environment through pre-determined percept and action APIs. The teamwork situation can be visualized and recorded with a stand-alone Monitor program (Fig. 2) for further analysis.
Fig. 2. The RoboCup simulation league is well-known for simulating multi-agent teamwork in playing soccer games. The teamwork situation in the virtual soccer field can be visualized and recorded with a stand-alone Monitor program.
B. System Architecture
RoboCup provides a visual environment for simulating multi-agent teamwork in soccer games. As shown in Fig 1(b), the RoboCup package uses a stand-alone Monitor program to read and display inner states of the environment, i.e., the passing of time, the current scores and the positions of the ball and the players. By delegating visualization functions to the Monitor, the Mediate Server can concentrate on responding the query-and-set requests for the agents, and updating the inner states of the virtual soccer field. The agent players use specific subroutines to transmit their requests. For instance, myposdir() returns the player's current position and its facing direction; sendTurn(), sendDash() and sendKick() send actions to the server for making a turn, a move and a kick respectively. The RoboCup simulation package thus does not limit the models of the multi-agent teamwork. The agent system can be implemented in any of the above collaboration architectures.
The RoboCup package architecture has a smart design that successfully simulates multi-agent teamwork in soccer games. However, the visualization effect can be improved if interactive environments are available for researchers to perform human-involved training. Human interaction not only enables supervised learning for the agents, but also supports researchers to repeat specific situations to thoroughly examine the agent behavior. Therefore, I propose a system developing tool in this project to develop ¡§interactive visual environments¡¨, including an enhanced RoboCup package, for Simulating and Training Collaborative Multi-agent Systems.
The proposed system development tool is a unique IDE (Integrated Development Environment). Besides the standard operations to generate, edit and compile user projects, the IDE offers advanced functions performed with the embedded Graphic Editor, Code Generator, and Library for Concurrent Collaboration, as illustrated in Fig 3. The¡§Graphic Editor¡¨ creates and edits the symbols for use in the simulation environments. The symbols can be categorized into four classes according to their properties and methods: (1) visible static symbols, such as the lines in a soccer or the players; (3) invisible static symbols, such as the such as the timer in sports games or the wind shear in virtual the graphic symbols into programmable object codes, develop the Monitor and the Mediate Server. The Frame users to develop individual agents and coaches. The ¡§ may exist in the STEAM or PTS models as an invisible predetermined conditions, can use coaches to manage strategy. The frame code of a coach is generated particularly, has communicates fully with all the other agents.
The library for concurrent collaboration is provided for human-based operation, not for inter-agent communication. In our ShareTone CSCW project, a service package called ¡§Collabench¡¨ was developed for collaborative event broadcasting and concurrent action. Collabench uses an event-driven framework to maintain a consistent state among replicated applications. A user intending to develop an environment that supports collaboratively human-based operation can include the Collabench library as a program segment of the Monitor. Significantly, the other parts of the created MAS project, i.e., the Mediate Server, the agents and the coach, are not influenced in this modulized system architecture. Replicated Monitors can be launched in connection to the Mediate Server, and multiple users are allowed to manipulate multiple agents concurrently. Unlike in on-line games, the concurrent manipulation is not restricted to the one-to-one mode: even if only two users are opening the Monitors, they can collaboratively operate ten agent players in a game. For further details about the development issues and prospective benefit, please refer to my ¡§Research Plan with Regard to the Application for the Position of Assistant Professor¡¨.
Fig. 3. Architecture of the proposed system development tool. It is a unique IDE (Integrated Development Environment) utility. Besides the standard operations to generate, edit and compile user projects, the IDE offers advanced functions performed with the embedded Graphic Editor, Code Generator, and Library for Concurrent Collaboration.
 R. C. Arkin and J. Diaz. Line-of-sight constrained exploration for reactive multiagent robotic teams. In Proc. 7th International Workshop on Advanced Motion Control, pages 455¡V461, 2002.
 P. Busetta and K. Ramamohanarao. An architecture for mobile BDI agents. In Proc. 1998 ACM symposium on Applied Computing, pages 445¡V452, 1998.
 P. R. Cohen and H. J. Levesque. Teamwork. NOU´S, 35:487¡V512, 1991.
 D. Emigh and I. Zaslavsky. Scientific visualization in the classroom. In Proc. 1998 ACM/IEEE conference on Supercomputing (CDROM), pages 1¡V7, 1998.
 Y. Guo and J. P. Muller. Multiagent collaborative learning for distributed business systems. aamas, 03:1156¡V1163, 2004.
 T. Hampel and R. Keil-Slawik. sTeam - designing an integrative infrastructure for web-based computer-supported cooperative learning. In Proc. 10th international conference on World Wide Web, pages 76¡V85, 2001.
 N. Houari and B. H. Far. Building collaborative intelligent agents: revealing main pillars. In Proc. 2005 Annual Conference on Electrical and Computer Engineering, pages 992¡V995, 2005.
 A. Howard. Multi-robot mapping using manifold representations. In Proc. 2004 IEEE International Conference on Robotics and Automation, volume 4, pages 4198¡V4203, New Orleans, LA, April 2004.
 K.-M. Jaa-Aro. Reconsidering the avatar : From user mirror to interaction locus, Ph. D. Dissertation, KTH, Numerisk analys och datalogi, Feb. 2004.
 J. R. Kok and N. A. Vlassis. Collaborative multiagent reinforcement learning by payoff propagation. Journal of Machine Learning Research, 7:1789¡V1828, 2006.
 J. Liu and J. Wu. Multiagent Robotic Systems. CRC Press, Inc., Boca Raton, FL, USA, 2001.
 R.-L. Liu. Collaborative multiagent adaptation for business environmental scanning through the internet. Applied Intelligence, 20(2):119¡V 133, 2004.
 T.-M. Rhyne, M. Tory, T. Munzner, M. Ward, C. Johnson, and D. H. Laidlaw. Information and scientific visualization: Separate but equal or happy together at last. In Proc. 14th IEEE Visualization 2003 (VIS'03), page 115, 2003.
 P. Stone and M. Veloso. Task decomposition, dynamic role assignment, and low-bandwidth communication for real-time strategic teamwork. Artificial Intelligence, 1999.
 L. Tijerina, J. J. Jackson, D. Pomerleau, R. A. Romano, and A. D. Pertersen. Driving simulator tests of lane separture collision avoidance systems. In ITS America Sixth Annual Meeting, April 1996.
 M. B. Vieira, A. de Albuquerque Ara´ujo, S. Philipp-Foliguet, M. Cord, M. Jordan, and J. Grillo. Navegador3D: An internet based flight simulator of urban centers. In 15th SIBGRAPI, page 432, 2002.
 N. Vlassis. A concise introduction to multiagent systems and distributed AI. Informatics Institute, University of Amsterdam, Sept. 2003.
 T. Vu, J. Go, G. Kaminka, M. Veloso, and B. Browning. MONAD: a flexible architecture for multi-agent control. In Proc. 2nd international joint conference on Autonomous agents and multiagent systems, pages 449¡V456, 2003.
 J. D. Wei and D. T. Lee. Priority-based genetic local search and it's application to the traveling salesman problem. Lecture Notes in Computer Science 4274, pages 424¡V432, 2006.