COBRA - COllective Behavior based on Realistic Aspects of human mobility

Mobility models have evolved from simply replicating statistical features using random models (e.g., random walk, random direction) to more sophisticated ones that attempt to demonstrate realistic human behavior and mobility patterns. Earlier models were expected to be simple, scalable, effortlessly configurable, and mathematically tractable. More recent models are data-driven and are able to generate synthetic traces that are comparable to real measurements. Also, models should demonstrate protocol performance similar to that under realistic conditions. Protocols such as Profile-cast and Bubble-rap harness the underlying structural dynamics of human communal behavior to transmit messages. For that purpose, models generated traces should reflect such dynamics that we evaluate using community detection. Previous studies have shown that the vast majority of mobility models are inadequate to depict realistic patterns. For example, using contact graphs authors have shown that models fail to capture bridging links between communities. Earlier, authors have shown that even carefully crafted models surprisingly result in structural dynamics and protocol performance that dramatically deviate from reality. The authors have observed that inter-contact patterns in theme parks are best described by the gamma distribution. These findings strongly motivate the need to re-visit mobility modeling to depict accurate human behavioral characteristics and communication network protocol performance.

A. Design description of COBRA

In this section, we describe the design of COBRA in more detail. The block diagram of COBRA is shown in the adjacent figure. The model components involve time structure and pause time, visitation location, epoch length distribution, event and mobility generator, and a trace generator. The model provides flexibility to independently configure each node’s Spatio-temporal patterns, thereby capturing the heterogeneous behavioral pattern and mobility at will. This makes COBRA distinct from other models and helps to capture the richness otherwise evident only in real measurements. We start with location visitation patterns of nodes that are the probability distribution of frequencies of their visits to a set of locations. This approach helps to capture skewed (heavily visited) as well infrequently visited locations. For example, a mobile user regularly goes to the office, but once in a while (say weekend) goes to the grocery store. In that sense, the probability to visit the office is much higher than stores. In the simulation setting, each location is a square geographical area (cell) with a constant edge length. Next, the duration of time a node spends in moving to a location is defined by epoch length. It starts from the endpoint of the previous location’s epoch and is generated from an exponential distribution equal to the size of the location. The offline behavior of the node is thus defined as the travel time from one epoch to another, measured through a speed and a direction (angle) movement for the chosen location. A roaming epoch is also defined when a node roams around the whole simulation area during some epoch, by assigning an additional location that corresponds to the whole simulation. Basically, a node chooses a new location probability and epoch and continues to move in that direction with a chosen speed. After each epoch, the node remains stationary in that location for the pause time drawn from the distribution. As evident from both the figures, there are few locations that are frequently visited with large pause times. In addition to that, we also gather periodicity, which provides flexibility to create multiple time periods with different locations and variable settings. The time periods are essentially the periodicity that is present in human mobility. For example, a weekly periodicity can be going to work during the weekdays and spending weekends at home or attending classroom lectures three times a week, etc. The epoch lengths and pause time, therefore, depend on the time periodicity. The model is data-driven and generates synthetic traces that can be compared against real measurements using metrics discussed in the framework section. We model time-dependent location selection process through Markov chains that maintain the spatio-temporal heterogeneity of individual nodes in the simulation area.

Files for Downloads

1. COBRA-Analysis.jar - Java source code that helps to perform the following activities
   1. Sanitize and process Real as well as COBRA generated measurements for future analysis (based on MIDAS)
   2. Used for generating SVD and perform similarity analysis of the mobile users (create spatio-temporal association matrix etc.).
   3. Gephi / Graphml and R File generation for social network analysis
   4. Encounter statistics (Encounter frequency, Encounter duration etc.). It also generates encounter frequency and inter encounter time distribution
   5. Spatio-temporal preferences
   6. ONE simulator files for running protocol analysis (Encounter based forwarding and Community based forwarding). Converting the traces for ONE simulator
   7. Input data file for running SMOOTH and its configuration parameters
   8. Matlab input file for the analysis of hierarchical clustering.