Adaptive classifier system-based dead reckoning

Abstract

Most dead reckoning implementations are based on DIS' specifications and only use a single prediction model during the whole simulation. However, several studies manage to improve dead reckoning's performance by defining prediction model selection strategies. Nevertheless, these approaches are either too generic and based on empirical results or too specific and only have few fields of application. This paper presents our approach that is meant to determine, among a set of extrapolation models, the one to apply in any given situation. It is based on classifier systems, adaptive evolutionary systems that are more generally involved to create artificial creatures in the field of "artificial life". Using such systems enable us to define a general model that can generate simulation-specific rules with relatively little work. Indeed, they just require defining the parameters that have to be taken into account and the criteria to optimize (e.g. accuracy, amount of updates...). Then, the system makes a set of rules emerge through a trial/error process in order to define more efficient and finer prediction model selection strategies.