Download presentation
Presentation is loading. Please wait.
Published byJasmin Cook Modified over 9 years ago
1
The Cat and The Mouse -- The Case of Mobile Sensors and Targets David K. Y. Yau Lab for Advanced Network Systems Dept of Computer Science Purdue University (Joint work with J. C. Chin, Y. Dong, and W. K. Hon) David K. Y. Yau Lab for Advanced Network Systems Dept of Computer Science Purdue University (Joint work with J. C. Chin, Y. Dong, and W. K. Hon)
2
Project Background Sensor-cyber project in national defense Near real-time detection, tracking, and analysis of plumes (nuclear, chemical, biological, …) Multi-university partnership funded by Oak Ridge National Lab Sensor testbed design and implementation Research team: Purdue, UIUC, LSU, U of Florida, Syracuse Personel Purdue: Jren-Chit Chin, Yu Dong, David Yau, Wing-Kai Hon Oak Ridge National Lab: Nageswara Rao Partnership with SensorNet initiative Sensor-cyber project in national defense Near real-time detection, tracking, and analysis of plumes (nuclear, chemical, biological, …) Multi-university partnership funded by Oak Ridge National Lab Sensor testbed design and implementation Research team: Purdue, UIUC, LSU, U of Florida, Syracuse Personel Purdue: Jren-Chit Chin, Yu Dong, David Yau, Wing-Kai Hon Oak Ridge National Lab: Nageswara Rao Partnership with SensorNet initiative
3
SensorNet Initiative Building comprehensive incident management system Coordinate knowledge and response effectively Provide data highway for processing sensor data Deliver near-real-time information for effective counter-measure Building comprehensive incident management system Coordinate knowledge and response effectively Provide data highway for processing sensor data Deliver near-real-time information for effective counter-measure Analysis, modeling and prediction Chemical Radiation Biological
4
Why Mobile? The mouse Evasion of detection Nature of “mission” The cat Improved coverage with fewer sensors Robustness against contingencies Planned or random movement (randomness useful) The mouse Evasion of detection Nature of “mission” The cat Improved coverage with fewer sensors Robustness against contingencies Planned or random movement (randomness useful)
5
Scenario—UAV Surveillance UAV detect radioactive plume Estimate position of plume source Control center predicts movement Emergency response
6
Mobility Model Four-tuple N: network area M: accessibility constraints -- the “map” T: trip selection R: route selection Random waypoint model is a special case Null accessibility constraints Uniform random trip selection Cartesian straight line route selection Four-tuple N: network area M: accessibility constraints -- the “map” T: trip selection R: route selection Random waypoint model is a special case Null accessibility constraints Uniform random trip selection Cartesian straight line route selection
7
Problem Formulation Two player game Payoff is time until detection (zero sum) Cat plays detection strategy Stochastic, characterized by per-cell presence probabilities Mouse plays evasion strategy Knows statistical process of cat’s movement, but not necessarily exact routes (exact positions at given times) Two player game Payoff is time until detection (zero sum) Cat plays detection strategy Stochastic, characterized by per-cell presence probabilities Mouse plays evasion strategy Knows statistical process of cat’s movement, but not necessarily exact routes (exact positions at given times)
8
Best Mouse Play Cat’s presence matrix given Network region divided into 2D cells P i,j gives probability for mouse to find cat in cell (i, j) Expected detection time “long” compared with trip from point A to point B Dynamic programming solution to maximize detection time Local greedy strategy does not always work Cat’s presence matrix given Network region divided into 2D cells P i,j gives probability for mouse to find cat in cell (i, j) Expected detection time “long” compared with trip from point A to point B Dynamic programming solution to maximize detection time Local greedy strategy does not always work
9
Optimal Escape Path Formulation For each cell j, mouse decides whether to stay or to move to a neighbor cell (and which one) If stay, expected max time until detection is E j [T stay ] If move to neighbor cell k, expected max time until detection is E j [T move(k) ] For cell j, expected max time until detection, E j [T], is largest of E j [T stay ] and E j [T move(k) ] for each neighbor cell k of j E j [T stay ] determined by cat’s presence matrix and expected cat’s sojourn time in each cell Optimal escape path is sequence of safest neighbors to move to, until mouse decides to stay How to compute E j [T] for each cell j? For each cell j, mouse decides whether to stay or to move to a neighbor cell (and which one) If stay, expected max time until detection is E j [T stay ] If move to neighbor cell k, expected max time until detection is E j [T move(k) ] For cell j, expected max time until detection, E j [T], is largest of E j [T stay ] and E j [T move(k) ] for each neighbor cell k of j E j [T stay ] determined by cat’s presence matrix and expected cat’s sojourn time in each cell Optimal escape path is sequence of safest neighbors to move to, until mouse decides to stay How to compute E j [T] for each cell j?
10
Computing E j [T] Initialize E j [T] as E j [T stay ] Insert all the cells into heap sorted by decreasing E j [T] Delete root cell 0 from heap For each neighbor cell k of 0, update E k [T] as E k [T] := max(E k [T], E k [T move(0) ]) Reorder heap in decreasing E j [T] order Repeat until heap becomes empty Initialize E j [T] as E j [T stay ] Insert all the cells into heap sorted by decreasing E j [T] Delete root cell 0 from heap For each neighbor cell k of 0, update E k [T] as E k [T] := max(E k [T], E k [T move(0) ]) Reorder heap in decreasing E j [T] order Repeat until heap becomes empty
11
Example Optimal Paths 0.0070.0090.010.0090.007 0.0090.050.10.050.009 0.010.10.080.10.01 0.0090.050.10.050.009 0.00750.0090.010.0090.0075 Path when mouse moves slowly Path when mouse moves quickly
12
Comparison with Local Greedy Strategy 0.0075 0.1 0.0075 0.10.080.10.0075 0.1 0.0075 Local greedy strategy: mouse will stay Dynamic programming strategy: mouse moves to cell with small probability of cat’s presence (0.0075) Current mouse position
13
If Cat Plays Random Waypoint Strategy Highest presence probability at the center of the network area Lowest presence probabilities at the corners and perimeters Good “safe havens” for mouse to hide Sum of presence probabilities is one n cats sum of probabilities n Equality for disjoint cats’ surveillance areas Highest presence probability at the center of the network area Lowest presence probabilities at the corners and perimeters Good “safe havens” for mouse to hide Sum of presence probabilities is one n cats sum of probabilities n Equality for disjoint cats’ surveillance areas
14
Distribution of Movement Direction in 150 m by 150 m Network Area
15
Cat’s Presence Matrix for Random Waypoint Movement
16
Cat’s Presence Matrix in 500 500 m Network for Random Waypoint Movement
17
Distribution of Movement Direction (a) Calculated probabilities of sensor moving towards the center cell from different current cells (b) Measured probabilities of sensor moving towards the center cell from different current cells
18
Analytical Cell Coverage Statistics (a) Expected number of trips before covering a cell (average = 11.431, maximum = 18.667) (b) Expected time before covering a cell (average = 59.604 s, maximum = 97.353 s)
19
Measured Cell Coverage Statistics (b) Expected number of trips before covering a cell (average = 10.301, maximum = 20.482) (b) Expected time before covering a cell (average = 52.721 s, maximum = 105.169 s)
20
Optimal Cat Strategy Maximize minimum presence probability among all the cells Eliminate safe haven Achieved by equal presence probabilities in each cell Will lead to Nash Equilibrium Zero sum game Pareto optimality Maximize minimum presence probability among all the cells Eliminate safe haven Achieved by equal presence probabilities in each cell Will lead to Nash Equilibrium Zero sum game Pareto optimality
21
Presence Matrices Random Waypoint Model Bouncing
22
Seeing Mouse Strategy
23
Blind Mouse Strategies Compared Average Detection Time Cat Strategies ScanBouncingRWP Mouse Strategies DP1083.26628.662823.26 RWP415.31442.23271.73 Stay511.50305.03226.13 V c = 10 m/s, V m = 10 m/s, R c = 25 m, R m = 0 m
24
Seeing Mouse Strategies Compared Average Detection Time Cat Strategies BouncingRWP Mouse Strategies Bouncing149.531455.28 Centric340.851092.29 Static92.39899.07 Stay10.2321.99 V c = 10 m/s, V m = 10 m/s, R c = 5 m, R m = 10 m
25
Effect of Mouse Sensing Range
26
Effect of Mouse Speed
27
Effect of Number of Cats
28
Minimum Sensing Range for Expected Random Waypoint Coverage Stationary mouse; cat in random waypoint movement Expected coverage desired by given deadline What is minimum sensing distance required? Stochastic analysis of shortest distance between cat and mouse within deadline Stationary mouse; cat in random waypoint movement Expected coverage desired by given deadline What is minimum sensing distance required? Stochastic analysis of shortest distance between cat and mouse within deadline
29
Lower Bound Cat-mouse Distance Network divided into m by n cells; each has fixed size s by s D(i, j): Euclidean distance between cell i and cell j N sets of cells sorted by set’s distance to mouse Each set of cells denoted as S j, 0 ≤ j ≤ N - 1 Each cell in S j is equidistant from the mouse; distance is D S j Distances sorted in increasing order; i.e., D S j < D S j+1 Network divided into m by n cells; each has fixed size s by s D(i, j): Euclidean distance between cell i and cell j N sets of cells sorted by set’s distance to mouse Each set of cells denoted as S j, 0 ≤ j ≤ N - 1 Each cell in S j is equidistant from the mouse; distance is D S j Distances sorted in increasing order; i.e., D S j < D S j+1
30
Example Equidistant Sets of Cells Mouse located at center of network area
31
Correlation between Cells Visited P i : probability that cat may visit cell i P S j : probability that cat may visit any cell in set S j P i : probability that cat may visit cell i P S j : probability that cat may visit any cell in set S j
32
Shortest Distance Probability Matrix from Cell i to Cell j 3-D probability matrix B Each element b i,j gives cat’s shortest distance distribution from mouse after trip from cell i to j is a size N vector: b i,j [k] is the probability that the shortest distance during the trip is D S k
33
where is the probability that D S 0 is shortest distance for trip l, and is probability that D S n is shortest distance for the trip, 1 ≤ n ≤ N - 1 Let denote, then is calculated as: Shortest Distance Probability Matrix after l Trips B l is the shortest distance probability matrix after l trips Computed by * operator B l is the shortest distance probability matrix after l trips Computed by * operator Each element of B l is calculated as:
34
Expected shortest distance The expected shortest distance between cat and mouse after l trips:
35
Approximate Expected Shortest Distance Approximate expected shortest distance from mouse after cat has visited k cells: P D j (k) is probability that after visiting k cells, a cell in S j is visited, but no cell in S i, i< j, is visited Approximate expected shortest distance from mouse after cat has visited k cells: P D j (k) is probability that after visiting k cells, a cell in S j is visited, but no cell in S i, i< j, is visited
36
Lower Bound Cat-mouse Distance for Random Waypoint Model (a) Expected speed = 5 m/s (b) Expected speed = 10 m/s(c) Expected speed = 25 m/s
37
Lower Bound Cat-mouse Distance for Indiana Map-based Model
38
Conclusions Considered cat and mouse game between mobile sensors and mobile target For random waypoint model, other coverage properties can be obtained analytically Expected cell sojourn time, expected time to cover general AOI, number of sensors to achieve coverage by given deadline, … Considered cat and mouse game between mobile sensors and mobile target For random waypoint model, other coverage properties can be obtained analytically Expected cell sojourn time, expected time to cover general AOI, number of sensors to achieve coverage by given deadline, …
39
Conclusions (cont’d) Many extensions possible Explicit account for plume explosion / dispersion models Model for sensor (un)reliability, interference, etc Explicit quantification of sensing uncertainty and its reduction Many extensions possible Explicit account for plume explosion / dispersion models Model for sensor (un)reliability, interference, etc Explicit quantification of sensing uncertainty and its reduction
40
Approximating the theoretical lower bound
41
Statistical properties of nodal movement Theorem 1 Let X and Y denote the width and height of the network, then the expected trip time is:
42
Statistical properties of nodal movement Theorem 2 The expected movement direction at a given position (x,y) can be expressed as
43
Statistical properties of nodal movement Theorem 3 Assume that each network cell is a circle with radius R. Within a single trip, the expected sojourn time (i)when the sensor is inside the starting (or ending) cell is (ii)when the sensor is not inside the starting or ending cell (i.e., when the sensor is crossing a cell in the trip) is Corollary 1 The expected cell sojourn time E[Ts] is Theorem 3 Assume that each network cell is a circle with radius R. Within a single trip, the expected sojourn time (i)when the sensor is inside the starting (or ending) cell is (ii)when the sensor is not inside the starting or ending cell (i.e., when the sensor is crossing a cell in the trip) is Corollary 1 The expected cell sojourn time E[Ts] is
44
Expected cell sojourn time Size (m) V max (m/s) V min (m/s) E[T i ] (s) E[T c ] (s) E[T max ] (s) T measure ( s) 301042.704.078.463.05 501044.176.2514.105.07 1001047.8211.7328.2110.11 50545.858.7714.107.65 501044.176.2514.105.07 501543.395.0914.103.96 501034.636.9518.815.81 501044.176.2514.105.07 501053.835.7511.284.56
45
Covering a target cell Theorem 4 (Naïve case) Assume each cell is equally likely to be covered next when a sensor leaves a cell. The expected number of cells visited by the sensor before covering a target cell equals the total number of cells in the network, i.e., E[k] = mn. Theorem 5 Let P i denote the probability that cell i is the next covered cell when the sensor leaves a cell. The expected number of cells visited by the sensor before covering the cell i, denoted as E[k i ], is 1/P i. The expected travel time before the mobile sensor covers the target is about: Theorem 4 (Naïve case) Assume each cell is equally likely to be covered next when a sensor leaves a cell. The expected number of cells visited by the sensor before covering a target cell equals the total number of cells in the network, i.e., E[k] = mn. Theorem 5 Let P i denote the probability that cell i is the next covered cell when the sensor leaves a cell. The expected number of cells visited by the sensor before covering the cell i, denoted as E[k i ], is 1/P i. The expected travel time before the mobile sensor covers the target is about:
46
Covering an Area-of-Interest (AOI) Theorem 6 Let P i be the probability for the target cell i to be covered by a sensor at any instance of time. The expected number of cells visited by the sensor (i)before covering all cells in AOI is (ii)before covering all cells in the network is Theorem 6 Let P i be the probability for the target cell i to be covered by a sensor at any instance of time. The expected number of cells visited by the sensor (i)before covering all cells in AOI is (ii)before covering all cells in the network is
47
How many sensors required? Theorem 7 Suppose that we are given the deadline D, the expected cell sojourn time E[Ts], and the probability Pi for the target cell i to be covered at any instance of time (if there is only one sensor). Then, to cover the target cell I with expected time smaller than D, the minimum number of sensor required is about.
48
Movement prediction result by using map based model Performance of LD with a map-based movement model within a 10,800 m by 11,500 m local area in Lafayette, IN, USA. (a) Communication distance vs. allowable delay (b) Distance savings vs. allowable delay. (c) Actual delay vs. allowable delay
49
Complexity of theoretical lower bound calculation Probability Matrix Approach: Update matrix for l trips: O((mn) 3 lN) Calculate E[d l min ]: O((mn) 2 N) Approximation Approach: O(mn+N) Probability Matrix Approach: Update matrix for l trips: O((mn) 3 lN) Calculate E[d l min ]: O((mn) 2 N) Approximation Approach: O(mn+N)
50
Comparison of two approaches (a) Expected speed = 5 m/s.(b) Expected speed = 10 m/s.(c) Expected speed = 25 m/s. Lower bound of expected communication distance (as a function of delay) computed by probability matrix approach and approximation approach. number of cells5 x 59 x 9 Time for probability matrix approach (s)0.17410.287 Time for approximation approach (s)0.0010.003 Comparison of computation time
51
Experimental results Comparison of the calculated theoretical lower bound of shortest communication distance to the target with the measurement results of the shortest distance, as well as the LD movement-prediction results. Average of one hundred 20,000-second runs Nodes moves in a 150 m by 150 m (or a 270 m by 270 m) network area, divided into 30 m by 30 m cells, Target is located in the center of network Modified random waypoint movement model With scenarios of expected speed of 5 m/s, 10 m/s, and 25 m/s. Comparison of the calculated theoretical lower bound of shortest communication distance to the target with the measurement results of the shortest distance, as well as the LD movement-prediction results. Average of one hundred 20,000-second runs Nodes moves in a 150 m by 150 m (or a 270 m by 270 m) network area, divided into 30 m by 30 m cells, Target is located in the center of network Modified random waypoint movement model With scenarios of expected speed of 5 m/s, 10 m/s, and 25 m/s.
52
LD performance (a) Expected speed = 5 m/s.(b) Expected speed = 10 m/s.(c) Expected speed = 25 m/s. Comparison of performance of LD with experimental expected shortest distance, and the theoretical lower bound in network area of size 150 m by 150 m divided into 25 cell. (Error bars are omitted because of small standard deviations).
53
LD performance (2) (a) Expected speed = 5 m/s.(b) Expected speed = 10 m/s.(c) Expected speed = 25 m/s. Comparison of performance of LD with experimental expected shortest distance, and the theoretical lower bound in network area of size 270 m by 270 m divided into 81 cell. (Error bars are omitted because of small standard deviations).
54
Effect of cell size (a) Expected speed = 5 m/s.(b) Expected speed = 10 m/s.(c) Expected speed = 25 m/s. Performance comparison of LD under different grid settings
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.