Information Based Sensor 5/22/2006 Management Information Based Sensor Management Ken Hintz Department of Electrical and Computer Engineering Center of Excellence in C 4 I Control/Tracking # 000000000 Item or Rev # 00000 Copy # 0000 Information Based Sensor DOCUMENT DATE: 5/22/2006 Page ___ of ___ 5/22/2006 Manager 1/23 Outline � Sensor manager vs. sensor scheduler � Information based sensor management � GMU SMS simulation/visualization demo � Summary DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 2/23 1
Information Based Sensor 5/22/2006 Management Motivation Who else is out there? Is anyone else Where should out there? I look next? What is the best way to Is anyone a find out? threat to me? DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 3/23 Sensor Management vs. Scheduling � Manager � Determine which observations sensors should make in order to best meet mission objectives � Scheduler � Determine the sequence of measurements to make within the constraints of sensor and platform capabilities DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 4/23 2
Information Based Sensor 5/22/2006 Management Need for a Sensor Manager � Cannot sense from all directions at the same time � Sensors are constrained in measurement or computation space � Tradeoff between accuracy and timeliness of measurements � Single sensor does not have “big picture” � Need to use integrated world model to direct individual sensor actions � Context of measurements defines contribution to mission DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 5/23 Predominant Sensor Management Model � Rule/doctrine/expert system approach � Rules are predicated on experience from last conflict � Rules are based on what experts think the next conflict will be like � Sensors and rules are based on sensor capabilities and the expected load � Rules are hierarchical and dogmatic � Global optimization is considered computationally intractable DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 6/23 3
Information Based Sensor 5/22/2006 Management Information Based Sensor Management � Shannon information � Information rate through a signal-to-noise and bandwidth limited channel based on coding � Used to develop methods for encoding data with no regard for content or meaning of data � Information based Sensor Management � Maximizes information gain to minimize valued uncertainty of platform’s world model by choosing � What to measure, � The sensor action to utilize, while � Leaving detailed sensor scheduling to individual sensors DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 7/23 Why Information? � Information, not data, is the raison d’etre for sensing � Need a common reference system within which to evaluate alternative sensing actions � Many performance measures for sensor systems are noncommensurate, e.g., P d , P kill , P lost_track , etc. � All sensing actions can be formulated as entropy changes, hence there is a computable information gain that can be associated with each � The amount of information (bits or nats) can be calculated independently of the sensor type, its characteristics, or which random variable one is interested in observing DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 8/23 4
Information Based Sensor 5/22/2006 Management Non-Shannon Information � Five types of sensor-based computable information about a process � Kinematic state � Search probability mass function � Target Identification � Cuer Information � Situation � All of these “informations” are based on the fundamental definition of information as being a measure of the change in uncertainty (entropy) about a random variable � Note that information is continually being lost about a random process and that information is a dynamic quantity DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 9/23 Types of Sensor “Informations” � Kinematic information � How much does a norm of error covariance matrix change as a result of a sensing action � Search information � How much does the probability mass function describing the location of targets change as a result of a sensing action � Identification Information � How much does the probability of a target being of a particular type change as a result of a sensing action � Cuer Information � How much does the detection of a target with one sensor type/mode change our ability to gain future information utilizing other sensors or modes � Situation Information � How much does our uncertainty about the intention or possible actions of a target change as a result of a sensing action DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 10/23 5
Information Based Sensor 5/22/2006 Management Situation Information � Purpose of a sensing system is to gather information in order to infer the intent of processes in its environment � Situation assessment is crucial to the sensor management paradigm since it allows us to decide what information we need � A modification of influence diagrams allows us to compute the expected gain in situation information which will result if we select one of several possible informations to request DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 11/23 SIEV Network � Modify influence diagram to produce situation information expected value (SIEV) network � Influence diagrams assign value to a decision based on expected value of utility function, essentially, expected monetary value � Interpret chance nodes in non- � Need probabilistic method of standard manner predicting the amount of situation � Compute utility based on information which can be gained topmost goals weighting of by taking one of several decisions information gain associated to acquire specific information only with situation chance independent of how we decide to nodes get that information DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 12/23 6
Information Based Sensor 5/22/2006 Management SIEV Net Nodes Chance nodes are subdivided into � Non-managed nodes � Sources of probabilistic data over which we have no control, e.g., air order of battle, electronic order of battle, are we being attacked?, etc. � Situation nodes � Hypotheses about our situation, e.g., hostile/friendly, target identification, target kinematics, etc . � Change in probability and/or error covariance of these nodes is situation information gain � Dynamically instantiated when target detected � (Sensor Manager) Managed nodes � Parameters associated with random variables whose values can be improved as a result of a sensor action DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 13/23 SIEV Net Nodes � Decision Node � Mutually exclusive decisions each of which equates to the launching of an information request � Utility node � Computes the gain in situation information associated with each possible information request � Based on changes in only the situation evidence nodes The information request is launched which produces the maximum SIEV DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 14/23 7
Information Based Sensor 5/22/2006 Management Annotated SIEV Net non-managed evidence nodes situation chance nodes managed evidence nodes Decision Node Topmost Goal Utility Values From GL Node DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 15/23 Information Is Not Enough � We know how to � Compute the amount of sensor information � Compute the amount of situation information � Need to incorporate dynamic mission goals into our valuation of a potential sensing action as expressed in a Goal Lattice � Topmost values related to situation information � Bottommost values related to value of sensing actions DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 16/23 8
Information Based Sensor 5/22/2006 Management Mission-Based GL Topmost Goals • Protect Self • Protect Friendlies • Conserve Power • Penetrate Defended Area • Collaborate Dynamic Goal Classes DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 17/23 Valued Information � For determining which situation information to request � Weight situation evidence nodes in the SIEV Net by topmost values of mission goal lattice � Produces mission-valued situation information expected value (SIEV) � For determining which particular sensing action to request to fulfill the information request � Evaluate applicable sensing functions as the product of the amount of information, the bottommost goal values, and the probability of obtaining that information to yield expected sensor information value (EIV) � Use EIV to determine which of several possible sensing actions to instantiate DOCUMENT DATE: 5/22/2006 5/22/2006 Information Based Sensor Manager 18/23 9
Recommend
More recommend
