ABSTRACT
In tune with the 24th Annual Cognitive Science Conference’s emphasis on application, this paper presents an empirical comparison between two methods used in agent tracking. The need to predict an agent’s intents or future actions has been well documented in multi-agent system’s literature and has been motivated by both systematically-practical and psychologically-principled concerns. However, little effort has focused on the comparison of predictive modeling techniques. This paper compares the performance of two predictive models both developed for the same, well-defined modeling task. Specifically, this paper compares the performance of a neural network based model and dead-reckoning model, both used to predict an agent’s trajectory and position. After introducing the background and motivation for the research, this paper reviews the form of the dead-reckoning algorithms, the architecture and training algorithms of the neural networks, the integration of the models into a large-scale simulation environment, and the means by which the performance measures are generated. Quantitative measures from our experiments indicate that, for the task considered, the neural network based model provides greater predictive utility, but at an increased cost in processing time. Performance measures are presented over increasing levels of error tolerance.
