Skip to main content

Optimization and machine learning are two major mathematical technologies for solving real problems in various industrial domains. The basic goal of optimization technologies is to effectively find the best allocation of limited resources, while machine learning technologies help us discover meaningful patterns in complicated data to predict the future for better outcome. Simulation is the third major approach, which we believe critical in analyzing complex systems.

One of the key ideas of ours is to use simulation technologies to support mathematical analysis of complex systems. In general, modeling complex systems in a fully analytical fashion is not easy, and simulation is often the only way to access the behavior of the system. Transportation systems, enterprises, and societies are typical example of complex systems, where many interacting micro-entities (agents) produce complex macro-level phenomena. We know that, for example, traffic congestion (that emerges from individual cars) and economic bubbles (that emerge from large numbers of transactions) are relatively common, but analyzing the emergence of such phenomena is still hard.

Agent-based simulation is a promising technology to analyze the complex systems. In our research, we are focusing mainly on following topics:

Development of Fundamental Technologies for Agent-based Simulations

When we investigate economic or traffic phenomena with agent-based simulations, the first challenge is the computational scalability. At IBM Research - Tokyo, we have a huge amount of accumulated research experience with agent-based software architectures. Our current software framework called ZASE (Zonal Agent-based Simulation Environment) supports ultra-large simulations that cannot be handled with simple implementations (in G. Yamamoto, H. Tai, and H. Mizuta, "A Platform for Massive Agent-based Simulation and its Evaluation," AAMAS 2007, pp. 900-902). ZASE provides a foundation for various simulation-based applications we are developing for such areas as traffic and marketing simulations. ZASE can efficiently manage enormous numbers of agents and express their dynamic behaviors even when multiple computers are required.

Most of the latest supercomputers use massively parallel architectures, where large number of multi-core processors and GPGPUs are interconnected. The massively parallel architectures have brought new challenges to simulator developers in handling large-scale agent-based simulations. Developers, who are often unfamiliar with parallel computing, have to be aware of the underlying multi-node architectures to take full advantage of their potential. In addition, they must pay attention to the difficult task of resource allocation to the individual nodes, which is especially critical in large-scale agent-based simulations with millions of agents. To comply with the massively parallel architectures, we developed XAXIS (X10-based Agents eXecutive Infrastructure for Simulation), a new platform for ultra-large agent-based simulations on massively parallel supercomputers. XAXIS is built on top of X10, a new parallel programming language suitable for multi-core architectures, which is being developed by IBM Research. In an ultra-large traffic simulation, our preliminary results showed up to a 34-fold speed-up on the TSUBAME 2.0 at the GSIC Center of the Tokyo Institute of Technology, a system with hundreds of compute nodes.

These techniques are also related to our research on the parallelization of analytics algorithms for large computer systems. In recent research, we devised a methodology for faster execution of analytics algorithms that need the huge amounts of computational power available on BlueGene/L, IBM's multi-node supercomputer (in T. Takahashi, and H. Mizuta, "Efficient Agent-based Simulation Framework for Multi-node Supercomputers", Winter Simulation Conference 2006, pp. 919-925). In this approach, the processing algorithm is divided into multiple modules and each module is dynamically relocated to an appropriate node to decrease execution time by using load monitoring of such data as the transmissions among nodes.

Multi-node distributed execution framework with dynamic load balancing
Fig. 1 - Multi-node distributed execution framework with dynamic load balancing (see, Takahashi and Mizuta, Proc. Winter Simulation Conference 2006, pp.919-925.)

Research on Industrial Applications of Simulation Technologies

Based on ZASE, we are implementing various simulators, such as markets and traffic systems.

Simulation Research for Marketing

Though marketing is key to business success, understanding mass media effects on sales is extremely difficult. Our first research in this area sought to understand how complex behaviors emerge in markets from the decisions made by a heterogeneous mix of human actors.

We were able to do detailed analysis of complex macro-market phenomena such as price bubbles using simple trading models with agent-based simulations (in H. Mizuta, K. Steiglitz, and E. Lirov, "Effects of Price Signal Choices on Market Stability", Journal of Economic Behavior and Organization (JEBO), Vol. 52, No. 2, pp. 235-251, 2003). Prof. Ken Steiglitz of Princeton University participated in our joint research using simulations and analysis to understand how various agents and various price signals affect stable markets.

We also performed simulations of the international emissions trading called for in the Kyoto Protocols. In this work, we were involved in pioneering research with a behavioral economics approach validated through simulation models and simulations games involving students from the University of Tokyo (in H. Mizuta and Y. Yamagata, "Transaction Cycle of Agents and Web-based Gaming Simulation for International Emissions Trading", Winter Simulation Conference 2002, pp. 801-806).

Now, we are doing research on forecasting the return on investment (ROI) of various marketing campaigns and advertisements. This is a challenge because the responses from consumers to marketing investments are nonlinear. To simulate such a non-linear responses, we model each consumer as an agent with a finite memory that selects actions using our special heuristics. These multi-agent simulations can efficiently evaluate the ROI based on the cumulative results of the individual nonlinear responses.


Research Themes


Research on Traffic Systems

Traffic phenomena are representative of complex systems that are interesting for both business standpoint and economic reasons. In this area, our research is integrating advanced data analytics technology and simulation technology in collaborations with universities and governments.

Here are some of the main research projects:

In these projects, we are developing a large-scale multi-agent traffic simulator called Megaffic (IBM Mega Traffic Simulator) that can handle large-scale fine-grained simulations of the traffic in a metropolitan area. This work involves a wide range of research topics from research on algorithms for real-time simulation and modeling of complex decision-making by humans to and data analytics for the trajectories of mobile objects. The integrated simulator will be able to assess emissions and provide sophisticated traffic information to support planning and evaluation of traffic policies.