1. Reconfiguration in Sensor Networks Qualifying Exam Karthik K Dantu Committee Gaurav Sukhatme (Chair) Ramesh Govindan Leana Golubchik Bhaskar Krishnamachari (Outside member) Cyrus Shahabi

2. 2/86 Sensor Actuator Networks Mars Exploration Marine monitoring Aerial Surveillance Forest canopy Volcano monitoring

3. 3/86 Sensor Actuator Networks: Challenges Robust connectivity in robot networks Distributing computation in tiered sensor networks

4. 4/86 Expected Contribution I.Spatial Reconfiguration A. Go from a connected network to a biconnected network using mobility B. Achieve route stability to perform topology control II.Functional Reconfiguration A. Reconfiguration of essential services in a sensor network B. Node functional reconfiguration

5. 5/86 Outline I.Spatial Reconfiguration A. Go from a connected network to a biconnected network using mobility B. Achieve route stability to perform topology control I.Functional Reconfiguration A. Reconfiguration of essential services in a sensor network B. Node-level functional reconfiguration

6. 6/86 Robot Networks Marine observationUrban search and rescue Formation controlCollaborative tasking Target tracking Robust connectivity between the robots!

7. 7/86 Articulation Point:Vertex or edge in a connected graph that connects two biconnected components. This edge or vertex is the only path between the two biconnected components Vertex articulation point Edge articulation point Biconnected Graph: Given a graph G(V, E), every vertex in V has atleast two vertex disjoint paths to every other node in V Terminology Biconnected Component: Subgraph of a given graph that is biconnected EA B I J F GC DH A BF C

8. 8/86 Robot Network Biconnectivity Problem Given a connected robot network, can they reassemble themselves with minimum movement to form a biconnected robot network?

9. 9/86 EA B I J F G C DH Achieving Biconnectivity

10. 10/86 Biconnectivity Algorithm: Introduction Conservative approach: Move one robot at a time Iterative: Improves connectivity of the network with iteration Each iteration potentially adds one new edge Only needs bearing measurement to neighbors Three phases in each iteration Compute biconnected components Identify potential neighbors Command movement

11. 11/86 Phase I: Compute biconnected components Tarjan’s algorithm - Centralized Depth-first search O(V+E) running time Each node sends its neighbor list to a central robot Central robot sends back component number Central robot notifies the articulation point R. Tarjan. Depth-first search and linear graph algorithms. SIAM Journal on Computing, 1(2):146–160, 1972.

12. 12/86 Phase II: At vertex articulation point Compute bearing to neighbors Sort neighbors based on relative bearing angle Identify neighbors belonging to different biconnected components and adjacent in this list Pick the pair with the least difference in relative bearing as candidate nodes E A B I J F GC D H

13. 13/86 Phase II: At edge articulation point Compute bearing to neighbors Sort neighbors based on relative bearing angle Identify the two nodes adjacent to the other edge articulation point on this list Exchange this information with the other edge articulation point Both mutually decide the nodes that are “closest” to each other Compute the expense of movement E A B I J F GC D H

14. 14/86 Phase III: Move command Node 1 Node 2Q1Q2Q3Q4 Q1 -a l - 45  135  - a h -a l 180  - a h 45  - a l a h - 135  a l - 270  90  - a h Q2 45  - a l 135  - a h 90  - a l 270  - a h 135  - a l 315  - a h Q3 135  - a l 315  - a h 180  - a l 360  - a h Q4225  - a l 415  - a h Adjacent Quadrants Same Quadrants

15. 15/86 Choosing a robot to move Compute expense of potential robots Expense of an AP is the sum of the expenses of the two robots it might move APs report cost to server Server chooses AP with least cost A B F C D E 2

16. 16/86 Avoiding losing edges Each moving robot keeps track of the signal strengths of its neighbors If RSSI falls below a threshold, it reports failure to the articulation point Articulation point commands the other node to move If both fail, AP reports failure to server Server sends move command to the AP that is next minimum cost If all APs report failure, algorithm terminates

17. 17/86 F G J G Biconnectivity algorithm: Discussion E A B I C DH Given a robot network B preassigned as server Nodes send neighbor lists to server B computes biconnected components and informs each node of its component number 1 2 3 Articulation points compute bearing of neighbors relative to themselves 180° 90° 270° 0° 30° 270° 315° 0° 315° J Articulation points compute the relative bearing to neighbors and choose the nodes to move

18. 18/86 Biconnectivity Algorithm: Simulation Simulations done using Player/Stage simulator Studied various randomly generated topologies (connected) Results shown for 16-node and 20-node networks Results averaged over 25 simulation trials actual computed Study error in Relative Bearing

19. 19/86 Biconnectivity Algorithm: Simulation Simulations done using Player/Stage simulator Studied various randomly generated topologies (connected) Results shown for 16-node and 20-node networks Results averaged over 25 simulation trials Study error in Relative Bearing Translation commanded executed

20. 20/86 Biconnectivity Algorithm: Simulation Simulations done using Player/Stage simulator Studied various randomly generated topologies (connected) Results shown for 16-node and 20-node networks Results averaged over 25 simulation trials commanded executed Study error in Relative bearing Translation Turn angle

21. 21/86 Biconnectivity: Simulation Sixteen node network BeforeAfter

22. 22/86 Robustness to Bearing Error AWGN noise with variance shown on x-axis Algorithm robust to large bearing error (upto 30°)!

23. 23/86 Robustness to Odometry Error AWGN Noise with variance a percentage of the command given

24. 24/86 Robustness to turn angle error AWGN Noise with variance as shown on x-axis

25. 25/86 Robustness to turn angle error AWGN Noise with variance as shown on x-axis

26. 26/86 Failure Modes Pinched Network

27. 27/86 Robot Experiments iRobot Create with e-boxes Antenna raised for better signal strength readings Use radio to compute relative bearing OLSR routing protocol Neighbor list obtained from routing table

28. 28/86 Robot Experiments - bearing computation Sample signal strength in the robot neighborhood according to pattern Measure 100 samples at each point Perform 3-D Principal Component Analysis Assume major axis as the direction of neighbor Experimented with wifi and zigbee radios Varied step size to study effect on bearing estimation Chosen step size 5 m Step size

29. 29/86 Bearing using RSSI - Outdoors

30. 30/86 Robot Experiments – 4-node Topology Before After

31. 31/86 Biconnectivity: 4-node Topology Initial Network Resultant Network

32. 32/86 Biconnectivity: 4-node Topology Trial 1Trial 2 Edge Actual angle Measured angle(T1) Bearing error (T1) Measured angle(T2) Bearing Error (T2) 3-4 -60  -45.7  14.3  -52.8  7.2  3-2 45  28.3  16.7  58.2  13.2  3-1 135  162.43  28.57  158.93  23.93  4-390  65.68  24.32  109.58  29.58  Average bearing error is 20.97°

33. 33/86 Before After Biconnectivity: 5-node Topology

34. 34/86 Biconnectivity: 5-node Topology Initial Network Resultant Network

35. 35/86 Biconnectivity: Summary Provided an algorithm that uses mobility to go from a connected graph to a biconnected graph using coarse relative bearing Demonstrated functionality of algorithm using simulation and experimentation Used radio signal strength for computing bearing

36. 36/86 Outline I.Spatial Reconfiguration A. Go from a connected network to a biconnected network using mobility B. Achieve route stability to perform topology control I.Functional Reconfiguration A. Reconfiguration of essential services in a sensor network B. Node-level functional reconfiguration

37. 37/86 Route Stability: Problem Control algorithms expect a stable graph to work on Routing protocol optimizes data forwarding path Path stability not a requirement for routing Frequent route switches confuse control algorithm Requirement unique to robot networks

38. 38/86 Route Stability: Example B A C D E F G

39. 39/86 Route Stability: Example B A C D E F G

40. 40/86 Route Stability: Example B A C D E F G

41. 41/86 Route Stability: Example B A C D E F G

42. 42/86 Route Stability: Example B A C D E F G

43. 43/86 Route Stability: Our solution Robot networks have information about direction of movement and possibly position This can be used to choose routes that are more likely to be stable Modify OLSR - a popular adhoc routing protocol

44. 44/86 Stability Metric: Direction Cue Direction Cue

45. 45/86 Stability Metric: Location and Direction Cue Stability metric is equal to the link duration

46. 46/86 OLSR Working Hello Messages to establish symmetric neighbors Exchange neighbor list to form list of 2-hop neighbors Identify Message Point Relays (MPR) to forward control messages

47. 47/86 OLSR: Data structure addition Link state routing protocol with optimizations for minimizing control messages Number of Information Bases for this purpose Multiple Interface Association Base Neighbor Information Base 2-Hop Neighbor Information Base MPR Information Base We add an information base for each cue – Direction Info base and Location Information Base

48. 48/86 OLSR: Message Modification Include velocity and position Add stability value along with every neighbor Hello Message Modification TC Message Modification Add stability value along with every neighbor

49. 49/86 OLSR: Processing Cues Hello Message Processing TC Message Processing Compute stability value and store Stability value of path is minimum of link stability value and the stability value announced in the tc message

50. 50/86 Modifying OLSR: Route computation New routes are picked based on the best stability value If a route exists, it is not modified unless the stability value falls below a threshold

51. 51/86 Simulation Setup Ns-2 with OLSR implementation from U Murcia, Spain Deploy robots in 500x500 area Mobility of robots is based on random waypoint mobility model Radius of communication is approximately 50 units – propagation model is Two Ray Ground Results are averaged over 10 trials each Percentage decrease in number of route switches is measured as the success metric

52. 52/86 Effectiveness – Dir Cue Only Route Stability vs. Density

53. 53/86 Effectiveness – Dir Cue Only Speed of travel vs. Route Stability

54. 54/86 Effectiveness – Dir + Loc Cue Density vs. Route Stability

55. 55/86 Effectiveness – Dir + Loc Cue Speed of travel vs. Route Stability

56. 56/86 Effect of error in location cue

57. 57/86 Effect of error in direction cue

58. 58/86 Route Stability - Summary Route stability – a unique problem in robot networks Simple direction cues yield up to 10% less route switches and up to 20% when used in combination with position cues Error in position is not very detrimental to the route stability unless the error is high Error in direction has a greater impact on route stability

59. 59/86 Related Work - Biconnectivity Mazda Ahmadi et al. Distributed algorithm for testing and maintaining biconnectivity in a robot network Poduri et al.Neighbor-Every-Theta graphs for k-connectivity Frederickson et al. Connectivity augmentation from k-1 connectivity to k-connectivity is NP-hard (k > 1) Kim et al.Optimization formulation for second-smallest eigen value of Laplacian matrix of the graph

60. 60/86 Related Work – Route Stability Ko and Vaidya Use location information to improve reactive routing in MANETs by using directional flooding Karp and KungGeographic routing in MANETs Ramachandra n et al. Measurement study of route stability in a MANET

61. 61/86 Outline I.Spatial Reconfiguration A. Go from a connected network to a biconnected network using mobility B. Achieve route stability to perform topology control I.Functional Reconfiguration A. Reconfiguration of essential services in a sensor network B. Node-level functional reconfiguration

62. 62/86 Motivation – Tiered Sensor Networks Mote-class sensor node Microserver node Site Link In-network data caching Internet Client Data Browsing and Processing Base Station Transit Network Patch Network Verification Network Sensor node Gateway

63. 63/86 Tiered Sensor Networks Microservers: More powerful processor and radio Tradeoff true distributedness for saving on communication Simple data collection from motes followed by computation at microserver

64. 64/86 Partitioning: The Basic Idea Motes Microservers

65. 65/86 Hierarchical Overlapped Coordination Consider an optimization problem of the form - subject to Where h and g are matrices and X

66. 66/86 Functional Dependence Table 1 0 1 …...... Columns correspond to the variables in the problem Rows correspond to constraints Table with entries such that (i,j)th entry is 1 if constraint i involves variable j, and zero otherwise

67. 67/86 Rearrange FDT.......... 00 0

68. 68/86 Problem reformulation subject to For each

69. 69/86 HOC Algorithm 1.Fix linking variables and solve problem by solving independent sub problems 2.Fix linking variables to their values determined in step 1 and solve problem by solving sub problems 3.Go to step 1 with fixed values of -linking variables determined by step 2 4.Repeat until convergence is achieved Nestor Michelena, Hyungju Park, Panos Papalambros and Devadatta Kulkarni, “Hierarchical Overlapping coordination under non-linear constraints”, American Institute of Aeronautics and Astronautics, 1998.

70. 70/86 Partitioning the Nodes Motes Microservers 1 1 11 1 1 1 1 1 1 1 1 1 2 22 2 2 2 2 222 2 2 2 1,2

71. 71/86 Connectivity-based Localization Let d i be the degree of node i Let (x i, y i ) be the initial location of node i Let R be the radius of communication For every neighbor node j, we formulate a constraint For every non-neighbor node k, we formulate a constraint

72. 72/86 SDP Relaxation Pratik Biswas, Yiyu Ye, “Semidefinite programming for ad hoc wireless sensor networks localization”, IPSN 2004.

73. 73/86 Simulation Setup Nodes distributed uniformly randomly in a square of side 100 units Radius of communication assumed to be 20 units Number of nodes varied from 100-250 in units of 50 Distributed simulations had four master nodes

74. 74/86 Centralized – Error vs. Degree

75. 75/86 Centralized - Error vs. Iterations

76. 76/86 Distributed – Error vs. Iterations

77. 77/86 Multiple masters

78. 78/86 Error – Edge Effects

79. 79/86 Incorporating other inputs Ranging Radio Interferometry N. Patwari and A.O. Hero,"Indirect Radio Interferometric Localization via Pairwise Distances," Third IEEE Conf on Embedded Sensor Networks (EmNets),

80. 80/86 Related Work WorkIdea Paek et al.

81. 81/86 Service reconfiguration - Summary Proposed a semi-centralized computational framework for data-fusion based services Trades off communication for computation Proof-of-concept evaluation with two services – localization and power-aware routing (described in document)

82. 82/86 Outline I.Spatial Reconfiguration A. Go from a connected network to a biconnected network using mobility B. Achieve route stability to perform topology control I.Functional Reconfiguration A. Reconfiguration of essential services in a sensor network B. Node-level functional reconfiguration

83. 83/86 Node Reconfiguration Sensor networks are deployed with preconfigured functionality as microservers or motes Functionality migration might benefit in multiple ways Learn event detection model of underlying application Provide requisite QoS Better network lifetime

84. 84/86 Related Work Power-aware base station positioning (Bogdanov et al 2004) Position base stations to maximize the data rate in a tree-based data collection model and fixed number of base stations Triage : Balancing Energy Consumption and Quality of Service in a Microserver (Banerjee et al 2007) Design surrogates for each function that delay using the microserver functionality while achieving a given QoS Sensor selection for predicting worst-case prediction error (Das et al 2007) Proposes algorithms to choose k sensors that would predict the sensed value with least chance or error Coverage control for mobile sensing networks (Cortes et al 2003) Proposes distributed gradient descent algorithms for optimal coverage using mobility

85. 85/86 LEAP-2 Node Microserver-class node Ability to measure per- component energy consumption uLeap – mote class node TinyOS 2 port ready Switch On/Off microserver based on requirement

86. 86/86 Timeline DateAgenda Jan-Mar 2009Node reconfiguration algorithm Apr 2009LEAP-2 Implementation May 2009Implement service reconfiguration June 2009 Experiments on node + service reconfiguration Implement location and direction cues on robot July 2009 Biconnectivity with stable routing – Experiments August-September 2009 Thesis writingOct-Nov 2009 DefenseDec 2009

87. 87/86 Publications Karthik Dantu and Gaurav S. Sukhatme. Connectivity vs. control: Using directional and positional cues to stabilize routing in robot networks. In submission, November 2008. 2. Karthik Dantu, Prakhar Goyal, and Gaurav S. Sukhatme. Relative bearing estimation from commodity radios. In Submission, September 2008. 3. Karthik Dantu, Prakhar Goyal, and Gaurav S. Sukhatme. Biconnected robot networks In Submission, August 2008. 4. Jesse Butterfield, Karthik Dantu, Brian P. Gerkey, Odest C. Jenkins, and Gaurav S. Sukhatme. Autonomous biconnected networks of mobile robots. In IEEE Workshop on Wireless Multihop Communications in Networked Robotics (WMCNR), Apr 2008. 5. Dustin McIntire, Timothy Chow, Karthik Dantu, Mansi Shah, Thanos Stathapoulos, Gaurav S. Sukhatme, William J. Kaiser The Low Power Energy Aware Processing (LEAP) Software Applications - Poster and Demonstration, In IPSN-SPOTS ’07: Proceedings of the 5th international conference on Information processing in sensor networks. 6. Karthik Dantu and Gaurav S. Sukhatme. Detecting and Tracking level sets of scalar fields using a robotic sensor network. In IEEE International Conference on Robotics and Automation (ICRA), pages 3665–3672, Apr 2007.

88. 88/86 Publications 7. Karthik Dantu and Gaurav S. Sukhatme. Rethinking data-fusion based services in tiered sensor networks. In IEEE Emnets ’06: Third Workshop on Embedded Networked Sensors,May 2006. 8. John Caffrey, Ramesh Govindan, Erik Johnson, Bhaskar Krishnamachari, Sami Masri, Gaurav S. Sukhatme, Krishna K. Chintalapudi, Karthik Dantu, Sumit Rangwala, Avinash Sridharan, Ning Xu, and Marco Zuniga. Networked sensing for structural health monitoring. In International Workshop of Structural Control Jun 2004. 9. Krishna Chintalapudi, Jeongyeup Paek, Omprakash Gnawali, Tat S. Fu, Karthik Dantu, John Caffrey, Ramesh Govindan, Erik Johnson, and Sami Masri. Structural damage detection and localization using netshm. In IPSN-SPOTS ’06: Proceedings of the 5th international conference on Information processing in sensor networks. 10. Karthik Dantu, Mohammad H. Rahimi, Hardik Shah, Sandeep Babel, Amit Dhariwal, and Gaurav S. Sukhatme. Robomote: enabling mobility in sensor networks. In IPSN-SPOTS ’05: Proceedings of the 5th international conference on Information processing in sensor networks. 11. Mohammad H. Rahimi, Rohit Mediratta, Karthik Dantu, and Gaurav S. Sukhatme. A testbed for experiments with sensor/actuator networks. Technical report, Institute for Robotics and Intelligent Systems Technical Report IRIS-02-417, 2002.

89. 89/86 Questions?

90. 90/86 Robustness to Bearing Error AWGN noise with variance shown on x-axis

91. 91/86 Robustness to Odometry Error

92. 92/86 RNB Decision Problem Given Robot Network Graph G(V,E) 2D Embedding of the network {(x i,y i |  i  V, x i, y i  R} Disc model of communication (radius r) One robot moves per time step (some distance d j ) RNB decision problem: Given a positive integer k, is there a sequence of movements M such that  d j ≤ k and resultant graph G M is biconnected?

93. 93/86 RNB Decision Problem Theorem: RNB Decision problem is NP-hard Planar Biconnectivity augmentation problem is NP-hard G. Kant and H. L. Bodlaender. Planar graph augmentation problems. In Workshop on Algorithms and Data Structures (WADS), pages 286– 298. Springer,1991. Given a planar weighted graph G(V, E) 3 4 1 7 6 3 10 11 5 A superset of weighted edges E’ NP-Hard to pick the minimum weight set of edges to make the graph biconnected

94. 94/86 Construction 1.Embed the planar graph in the plane [O(n)] 2.Let us assume that each vertex is a robot and each edge is a communication link 3.Let us further assume that each robot can only move in the direction of another node Norishige Chiba, Takao NishizekiShigenobu Abe Takao Ozawa, “A Linear Algorithm for Embedding Planar Graphs using PQ-Trees”, Journal of Computer and System Sciences, pp 54-76, Feb 1985. Theorem: RNB Decision problem is NP-hard Proof: Given Planar Biconnectivity Augmentation problem is NP-hard 3 10 11 5 4.Let the cost of moving towards the node to form an edge be the cost of augmenting the edge in the original problem Solving RNB problem solves the planar biconnectvitiy augmentation problem NP-Hard! RNB Decision Problem

95. 95/86 Problem Formulation If node j is a neighbor of i but the initial estimates are greater than R apart, then add the difference to the cost If node k is not a neighbor of i but the initial estimates are less than R apart, then add the difference to the cost

96. 96/86 Unconstrained Minimization Problem