Magazine article AI Magazine

SPADES: A System for Parallel-Agent, Discrete-Event Simulation

Magazine article AI Magazine

SPADES: A System for Parallel-Agent, Discrete-Event Simulation

Article excerpt

Main System Features

Simulations are an excellent tool for studying AI. They can allow the systematic modification of parameters of the environment, execute the large number of trials often required for machine learning, and facilitate the interaction of agents created by different research groups. On the one hand, many general simulation environments do not address the special concerns of the Al community, such as the computation time of the agent being an integral part of its behavior. On the other hand, many simulators created in the Al community fail to take advantage of the vast work in the simulation community for designing stable, repeatable, and efficient simulations.

This article discusses SPADES, the system for parallel-agent discrete-event simulation. (1) The system is designed to support simulations for the AI community without being tied to any particular simulated world. SPADES provides support for simulations with agents running in parallel across multiple machines and for the tracking of the computation time used by these agents. By taking advantage of work in discrete-event simulation, the middleware eases the design of a simulation by taking care of many of the system details required to handle distribution in an efficient and reproducible way.

Some primary features of the system are the following:

First is agent-based execution, including explicit support for modeling latencies in sensation, thinking, and acting. Agents are explicit, fundamental components of the system. The use of the word agent here refers to a software entity that has an explicit sense-think-act cycle interaction with the world. These components can overlap in time, and SPADES explicitly allows for this overlap.

Second, agents can be distributed among multiple machines. Distribution across machines can increase the pace of the simulation.

Third, the result of the simulation is unaffected by network delays or load variations among the machines. The efficiency of the simulation can be affected by these factors, but the results of the simulation are not. This is in sharp contrast to the SOCCER SERVER, the current simulator for the RoboCup simulation league (Noda et al. 1998).

Fourth, the architecture for the agents is unconstrained and does not require that the agents be written in a particular programming language. Unlike other systems that track the computation time used by agents (such as MESS [Anderson 1995], the only requirement on the agents is that they run as their own process on the machine. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.