Using SPC Techniques to Interpret the Input Data-Stream To An Adaptive-Agent Discrete-Event Simulation Element

摘要

This papmarizes an application of the common statistical process control tools, which are traditionally used to manage and control the variability of a production process. In this research this permits an adaptive-agent in a discrete simulation experiment to evaluate an input data stream, determine if there is a statistically significant shift in the behavior of the population that generated the data stream, and subsequently direct succeeding discrete simulation elements to process entities in different ways. Issues to be addressed in later studies are identified and form a list of future research goals. In a recent paper presented at the 1999 Advanced Technologies Simulation Conference, this author examined some of the salient design issues in discrete-event simulation models containing an adaptive-agent. Essentially in models containing adaptive-agents, the arrival of an entity signals the model to assign attributes to that entity which can be tracked throughout the system, where ordering and routing of that entity can be defined by the system model, and where statistics such as count, utilization, cycle time, etc. Can be compiled about the entity. The term agent refers to a system element that is designed to take action on its own, selecting its action using information it chooses from a range of available options. Adaptive-agent system elements create new system behaviors in non-random ways with the goal of incrementally achieving a better fit to the requirements of its environment. The focus of this paper is to examine statistical criteria that adaptive-agent elements might use to determine which, from the range of options, it should choose and in this way change the behavior of the entire simulation model. Traditional time series analysis provides a valuable classification scheme to describe the components of an input data stream. These are a base level, a trend, a seasonal, and random or noise components. A common application of elementary tools of linear regression analysis augmented with an assessment of seasonality is often used to analyze the data with each data point being of equal importance. Conceptually, the analysis of data streams by adaptive-agents must be thought of in a different context where recent data points are most important and the "value" of data points diminishes with age. Specifically, an adaptive-agent element is not necessarily interested in a "forecast" of the value of the next data point rather the adaptive-agent must determine if the most recent data points come from a different population {for example from a different season} and subsequently needs to change the behavior of the simulation model. In other words, if the adaptive-agent senses that inputs to the system are coming from population A, it will instruct the simulation model to process arrivals in a certain way, if however it senses that inputs are coming from population B, it will instruct the simulation model to process the arrivals in a different way. The fundamental purpose of statistical process control tools is to determine if a production process has changed and is in fact generating items from a statistically different population. The traditional tools of X-bar, R, np, and c charts are used to make these determinations. Using these tools, including the work of Nelson {1985}, provides a means to analyze the input data stream to an adaptive-agent simulation element and thus furnish a hierarchy for the adaptive-agent to determine if and how it should alter the behavior of the simulation model. This paper describes the application of these tools in a simple simulation model of a production system containing a single adaptive-agent and it discusses the implications of this research in more complex simulation models containing multiple adaptive-agent elements. The short-term contribution of this research project is that it extends the use of well-known and proven SPC analysis tools in a new application and in a new context. For the