Modeling How Thinking About the Past and Future Impacts Network Traffic with the GOSMR Architecture.

摘要

We present the GOSMR architecture, a modular agent architecture designed to actuate web browsers and other network applications, and demonstrate the importance of modeling how users think about the past and future in accurately modeling network traffic. The architecture separates the hierarchical generation of goals and incentives (Behaviors) from hierarchical implementations of their pursuit (Actions). Cognitive aspects modeled include the hyperbolic discounting of future payoffs, the chance a user forgets a task, and the ability of the user to defer tasks for later. The system also uses a logical grammar to allow agents to communicate Beliefs and delegate Actions. Using records of weekend Virtual Private Network traffic from over three thousand users at a medium-scale enterprise, we provide evidence for the importance of the forgetting, payoff discounting, and procrastinating aspects of the model, showing that agent payoff discounting and look ahead predict the observed spike in Sun- day night traffic, while forgetfulness can explain a decline in activity on Saturday where the utility of login should be increasing. We then use the learned parameters from this fitting to actuate agents visiting a social networking website hosted on a virtual machine, and we measure the impact of increasing or decreasing the perceived ease of login on the hourly volume of network traffic at peak times.