1. “Sociological Orbits” Mobility Profiling and Routing for Mobile Wireless Networks Hung Q. Ngo Computer Science & Engineering State University of New York at Buffalo

2. Jul 11, 2006 Acknowledgement On-going project with Prof. Chunming Qiao Joy Ghosh Duc Ha S. K. Yoon Dr. Sumesh Philip (former student)

3. Jul 11, 2006 Outline Mobility Impact on Routing Sociological ORBIT Mobility Framework Mobility Profiling Techniques and Applications A Fundamental Routing Problem on ICMAN Sociological Orbit aware Location Approximation and Routing (SOLAR) Conclusion

4. Jul 11, 2006 Mobility Impact on Routing Node Mobility  Dynamic network topology Proactive protocols (LS, DV) are inefficient  Need to exchange control packets too often  Leads to congestion Reactive protocols (DSR, LAR) are better suited, however  Locating a node incurs more delay  Route maintenance is tricky as nodes movetricky To strike a balance  need mobility modeling

5. Jul 11, 2006 Mobility Models in the Literature Random Waypoint, Weighted Random Waypoint  simple, but impractical!! Entity based  individual node movement  Jardosh et al., MOBICOM’03  Lin et al., INFOCOM’04 Group based  collective group movement  Hong et al., MSWIM’99, MCM’01 Scenario based  geographical constraints  Lam et al., IEEE Comm. Mag. 97  Markulidakis et al., IEEE Per. Comm. 97  Liu et al., IEEE JSAC 98

6. Jul 11, 2006 Advantages of Node Mobility – Individual node’s view of network

7. Jul 11, 2006 Advantages of Node Mobility – Node’s view of network through “acquaintances”

8. Jul 11, 2006 Impact of mobility on protocol performance F. Bai, N. Sadagopan, and A. Helmy, “Important: a framework to systematically analyze the impact of mobility on performance of routing protocols for adhoc networks”, Proceedings of IEEE INFOCOM '03, vol. 2, pp. 825-835, March 2003.

9. Jul 11, 2006 Our Motivations Observations MANET is often comprised of wireless devices carried by people living within societies Social activities impose constraints on user movements Steps to take Study the social influence on user mobility (e.g., realization of special regions of some social value) Identify a macro level (thus, lightweight) mobility profile per user Use this profile to aid macro level soft location management and routing

10. Jul 11, 2006 Mobile Users influenced by social routines visit a few “hubs” / places (outdoor/indoor) regularly “orbit” around (fine to coarse grained) hubs at several levels Sociological Orbit Framework

11. Jul 11, 2006 Illustration of A Random Orbit Model (Random Waypoint + Corridor Path) Conference Track 1 Conference Track 3 Cafeteria Lounge Conference Track 2 Conference Track 4 Posters Registration Exhibits

12. Jul 11, 2006 Random Orbit Model See Ghosh et al., Adhoc Networks Journal, 2005.

13. Jul 11, 2006 Hub Based Mobility Profiles and Prediction On any given day, a user may regularly visit a small number of “hubs” (e.g., locations A and B) Each mobility profile is a weighted list of hubs, where weight = hub visit probability (e.g., 70% A and 50% B) In any given period (e.g., week), a user may follow a few such “mobility profiles” (e.g., P1 and P2) Each profile is in turn associated with a (daily) probability (e.g., 60% P1 and 40% P2) Example: P1={A=0.7, B=0.5} and P2={B=0.9, C=0.6}  On an ordinary day, a user may go to locations A, B and C with the following probabilities, resp.: 0.42 (=0.6x0.7), 0.66 (= 0.6x0.5 + 0.4+0.9) and 0.24 (=0.4x0.6)  20% more accurate than simple visit-frequency based prediction  Knowing exactly which profile a user will follow on a given day can result in even more accurate prediction On any given day, a user may regularly visit a small number of “hubs” (e.g., locations A and B) Each mobility profile is a weighted list of hubs, where weight = hub visit probability (e.g., 70% A and 50% B) In any given period (e.g., week), a user may follow a few such “mobility profiles” (e.g., P1 and P2) Each profile is in turn associated with a (daily) probability (e.g., 60% P1 and 40% P2) Example: P1={A=0.7, B=0.5} and P2={B=0.9, C=0.6} On an ordinary day, a user may go to locations A, B & C with the following probabilities: 0.42 (=0.6x0.7), 0.66 (= 0.6x0.5 + 0.4+0.9), 0.24 (=0.4x0.6) 20% more accurate than simple visit-frequency based prediction Knowing exactly which profile a user will follow on a given day can result in even more accurate prediction

14. Jul 11, 2006 Traces Used Profiling techniques applied to ETH Zurich traces  Duration of 1 year from 4/1/04 till 3/31/05  13,620 wireless users, 391 APs, 43 buildings  Grouped users into 6 groups based on degree of activity  Selected one sample (most active) user from each group Mapped APs into buildings based on AP’s coordinates, and each building becomes a “hub”  Converted AP-based traces into hub-based traces Other traces  Expect similar results from Dartmouth’s traces  No sufficient AP location info from other traces  UMass’s traces are for buses, more predictable than users  Need to obtain actual users’ traces with GPS

15. Jul 11, 2006 Orbital Mobility Profiling Obtain each user’s daily hub lists as binary vectors Represent each hub list (binary vector) as a point in a n-dimensional space (n = total number of hubs) Cluster these points into multiple clusters, each with a mean  Using the Expectation-Maximization (EM) algorithm to train the model based on a Mixture of Bernoulli’s distribution  Probe other classification methods: Bayesian-Bernoulli’s Each cluster mean represents a mobility profile, described as a probabilistic hub visitation list User’s mobility is aptly modeled using a mixture of mobility profiles with certain “mixing proportions”

16. Jul 11, 2006 Profiling illustration Obtain daily hub stay durations Translate to binary hub visitation vectors Apply clustering algorithm to find mixture of profiles

17. Jul 11, 2006 Profile parameters for all sample users

18. Jul 11, 2006 Hub-based Location Predictions - I Unconditional Hub-visit Prediction  Prediction Error = Incorrect hubs predicted over Total hubs  SPE – Statistical based Prediction Error SPE-ALL: (n+1) th day prediction based on hub-visit frequency from day 1 through day n SPE-W7 : (n+1) th day prediction based on hub-visit frequency within last week, i.e., day (n-7) through day n  PPE – Profile based Prediction Error PPE-W7 : (n+1) th day prediction based on profiles of the last week, i.e., day (n-7) through day n  Prediction Improvement Ratio (PIR) PIR-ALL = (SPE-ALL – PPE-ALL) / SPE-ALL PIR-W7 = (SPE-W7 – PPE-W7) / SPE-W7

19. Jul 11, 2006 Unconditional Prediction Results The profile mixing proportions vary with every window of n days

20. Jul 11, 2006 Hub-based Location Predictions - II Conditional Hub-visit Prediction  Improvement given current profile is known/identifiable  It is possible sometimes to infer profile from current hub information alone  Our method effectively leverages information when available Sample user categories Target Hub ID: will the user visit this hub?The current day in questionPredicted probability using visit frequency Indicator (Current) HubCurrent ProfilePredicted probability based on profile Actually visited H t on day D or not

21. Jul 11, 2006 Hub-based Location Predictions - III Hub sequence prediction based on hub transitional probability Prediction Accuracy = 1 – (incorrect predictions / total predictions) Scenario 1: only starting hub is known for sequence prediction Scenario 2: hub prediction is corrected at every hub in sequence Better performance with increasing knowledge – intuitive Statistical based Prediction Accuracy (SPA) – no profile informationProfile based Prediction Accuracy (PPA) – no time informationTime based Prediction Accuracy (TPA) – temporal profiles

22. Jul 11, 2006 Applications of Orbital Mobility Profiles Location Predictions and Routing within MANET and ICMAN  We will discuss an example of routing on ICMAN  We have several other papers in this area (see website at the end) Anomaly based intrusion detection  unexpected movement (in time or space) sets off an alarm Customizable traffic alerts  alert only the individuals who might be affected by a specific traffic condition Targeted inspection  examine only the persons who have routinely visited specific regions Environmental/health monitoring  identify travelers who can relay data sensed at remote locations with no APs Others

23. Jul 11, 2006 Routing challenges in ICMAN ICMAN (Intermittently Connected MANET)  Features of DTN/ICN + MANET Lack of infrastructure and any central control May not have an end-to-end path from source to destination at any given point in time Conventional MANET routing strategies fail User mobility may not be deterministic or controllable Devices are constrained by power, memory, etc. Applications need to be delay/disruption tolerant

24. Jul 11, 2006 On Problem’s Complexity Basic model :  G = (V,E) be a directed graph  V = ICMAN users; E = probabilistic contact between users  Let A be a routing algorithm and G(A) be the delivery sub-graph induced by A subject to some constraint Basic tradeoff: overhead vs delivery probability Possible constraints to limit overhead  Constraint 1: each intermediate node forwards packets to at most k downstream neighbors  Constraint 2: G(A) has at most k edges

25. Jul 11, 2006 On Problem’s Complexity (cont’) The centralized version: given G and k, find a delivery subgraph H where Conn 2 (H) is maximized, subject to |E(H)| ≤ k Two Negative Theorems 1. Computing Conn 2 (H) is #P-complete 2. Finding H maximizing Conn 2 (H) is #P-hard What do we do now?  Approximate Conn 2 (H) by another poly-time computable function f, then find H that maximizes f(H)  Develop heuristics routing algorithms

26. Jul 11, 2006 Our Problem’s Setting Slightly different from the basic problem just discussed  Mobility profiles give contact probabilities  But, contact probabilities do not give mobility profiles We make use of the mobility profiles in our routing heuristics!

27. Jul 11, 2006 User level routing strategies Deliver packets to the destination itself Intermediate users store-carry-forward the packets Mobility profiles used to compute pair wise user contact probability P(u,v) via Markov Process Form weighted graph G with edge weights w(u,v) = log (1/P(u,v)) Apply modified Dijkstra’s on G to obtain k-shortest paths (KSP) with corresponding Delivery probability under following constraints  Paths are chosen in increasing order of total weights (i.e., minimum first)  Each path must have different next hop from source S-SOLAR-KSP (static) protocol  Source only stores set of unique next-hops on its KSP  Forwards only to max k users of the chosen set that come within radio range within time T D-SOLAR-KSP (dynamic) protocol  Source always considers the current set of neighbors  Forwards to max k users with higher delivery probability to destination

28. Jul 11, 2006 Hub level routing strategy Deliver packets to the hubs visited by destination Intermediate users store-carry-forward the packets Packet stored in a hub by other users staying in that hub (or using a fixed hub storage device if any) Mobility profiles used to obtain delivery probabilities (DP), not the visit probability, of a user to a given hub  i.e. user may either directly deliver to hub by traversing to the hub, or may pass onto other users who can deliver to the hub Fractional data delivered to each hub proportional to the probability of finding the destination in it Routing Strategy  SOLAR-HUB protocol

29. Jul 11, 2006 SOLAR-HUB Protocol P d n i h j : delivery probability (DP) of user n i to hub h j P t n i h j : probability of user n i to travel to hub h j h(n i ): hub that user n i is going to visit next P c n i n k (h j ): probability of contact between users n i & n j in hub h j N(n i ): neighbors of user n i P d n i h j = max(P t n i h j, max k (P c n i n k (h(n i ))*P t n k h j )) Source n s will pick n i as next hop to hub h j as:  {n i | max(P d n i h j ), n i Є N(n s )} iff P d n i h j > P d n s h j Packet Delivery Scheme  Source transmits up to k copies of message k/2 to neighbors with higher DP to “most visited” hub k/2 to neighbors with higher DP to “2nd most visited” hub  Downstream users forward up to k users with higher DP to the hub chosen by upstream node

30. Jul 11, 2006 Simulation Parameters for GloMoSim

31. Jul 11, 2006 Performance – Number of Hubs Overhead of EPIDEMIC is much more than others and had to be omitted from plot Overall D-SOLAR-KSP performs best

32. Jul 11, 2006 Performance – Number of Users Overhead of EPIDEMIC is much more than others and had to be omitted from plot Overall D-SOLAR-KSP performs best like before because it is the most opportunistic in forwarding to any of its current neighbors

33. Jul 11, 2006 Performance – Cache Size (Only SOLAR) All versions fair better with more cache Overall D-SOLAR-KSP performs best

34. Jul 11, 2006 Performance – Cache Timeout (Only SOLAR) All versions fair better with larger timeout Overall D-SOLAR-KSP performs best

35. Jul 11, 2006 Conclusion Mobility has severe impact on routing performance A practical mobility framework should the sociological influence on user movement into account Wireless users can be profiled based on their social activities Mobility profiles are useful not only for routing (e.g., SOLAR protocols) but also other applications such as location prediction, resource allocation, etc. SOLAR Project: (for more information)  http://www.cse.buffalo.edu/~joyghosh/solar.html

36. Thank You! Questions?