|
You are at: Wagner
Home > Technologies > Data Fusion Tracking
> Course of Action
Course of Action Assessment
and Creation
This research draws on classical Bayesian Inference combined
with powerful new tools called Quasi-Axiomatic Theories (QATTM) to provide a comprehensive and adaptive
situation analysis decision aid. The Bayesian Inference component
provides automatic assessment of hypotheses based on the continuous
arrival of positive (contact) and negative (search) information.
The QATTM component provides the adaptive
mechanism by which new hypotheses can be formed in the presence
of conflicting or seemingly erroneous data. A graphical user
interface allows the operator to interact with the adaptive algorithms
to create and assess hypotheses.
An adaptive system of algorithms promises to improve situation
assessment and reduce confusion in the demanding, time-critical
environment of the battlefield. These results can also be applied
to police surveillance, counter narcotics and industrial intelligence
analysis.
Contents
Course
of Action Assessment and the Bayesian Approach
The assessment of routes taken by the enemy is just one component
of the intelligence reporting process and, along with other types
of reports, can be used to infer enemy intentions.
In this simple example, we represent the enemy intentions
as alternate Courses of Actions (COAs) and all the possible sensor
reports as Events and all collection activities are termed "searches."
The following diagram illustrates the relationships between
COAs, events, searches, search results, and updated COA likelihood
assessments. The diagram shows four competing and mutually exclusive
COAs A, B, C, and D. These are not directly detectable because
they represent the enemy's closely held plans or intentions.
However, in carrying out their plans, the enemy must take some
observable actions, or events, that can be detected. In this
simple example, there are 10 events and each COA can trigger
one or more of them. Our assets conduct searches designed to
detect the events and we use those detections and non-detections
to improve our knowledge about the relative likelihoods of the
hypotheses.
The operator or commander has provided a likelihood assessment
for each of the arrows associating COAs with events and searches
with event detections. The commander also has provided prior
probability assessments of the COAs.
Hypothesis
Updating Based Over Time
How do you know when you don't know?
The following numerical results illustrate the operation of
the inference equations for this example using probabilities
for COAs, events, searches, and detections chosen at random,
including a NULL hypothesis (i.e., real COA has not been specified).
In addition, the occurrences of events were chosen at random,
as were the detections , over 14 time periods. The first bar
graph summarizes the information available to the inference equation.
There are many searches, or "looks," but only 3 detected
events. This represents a typical reconnaissance or surveillance
activity where the number of detections is small in comparison
to search effort.
The next graph shows how the Bayesian algorithm updates the
COA probabilities over time. In this example, we modeled COA
"D" as a NULL hypothesis, meaning that the algorithm
did not know anything about COA "D" in advance
(in particular, the expected distribution of events). Even so,
the algorithm was able to promote the NULL hypothesis probability
from a starting value of 0.2 to over 0.5 after the first two
detections at times 10 and 11, thus determining that the NULL
hypothesis was highly likely and the real COA was not in the
list.
This is a very simple example of the application of Bayesian
probability methods to the situation assessment problem. We have
developed a Realistic Scenario that combines the Monte Carlo
route model with the Bayesian updating process. In this scenario,
we have multiple units, NTC terrain, multiple COAs taken from
actual OPFOR plans, and we show how a simple lack of reporting
from forward units can quickly indicate the actual COA in progress.
Modeling
Enemy Units' Motion Over Terrain
Wagner Associates has developed a sophisticated search model
called SSPS, which we have enhanced under recent projects to
add the capability to model targets moving in terrrain. The Monte
Carlo method works by choosing large numbers of sample paths,
each of which follows the rules of motion but with its own statistical
variation. When new information is received, such as when a sensor
is activated at a certain location and does not see the target,
SSPS adjusts the relative probabilities of the sample paths to
change the over all probability distribution.
SSPS displays target location probability using "probability
maps" and can generate a sequence of maps to present a moving
picture of the target.
You can look at a sample of one of these moving probability
maps as individual movies. In the movie, the sequence repeats
first for the prior target model (no information) and then accounting
for negative information (look for the blue circles designating
sensor visibility). Watch the banner at the top of the frame
to see which unit and whether negative information is in effect
for the frame. Click here to view
the movie.
Positive
and Negative Information and COA Assessment
SSPS updates these probability maps by computing probability
of detection and modified weights for the individual motion models.
Given these data and the prior probabilities for COA, we can
now produce a modified COA probability distribution for any time
during the scenario.
In this example we use use all five COAs specified in the
original problem. As expected, the Bayesian assessment module
will show the probability of the tru COA rising rapidly as the
scenario unfolds.
The graphic shows the change in Course of Action probability
over the first three hours of the scenario. The color codes are:
| COA ONE - MOUNTAIN PASS NORTH |
 |
| COA TWO - MOUNTAIN PASS NORTH |
 |
| COA THREE - NORTH VALLEY |
 |
| COA FOUR - CENTRAL VALLEY |
 |
| COA FIVE - SOUTHERN ROUTE |
 |
| NULL COA |
 |
Now we watch the scenario unfold. The thickness of the color
indicates the relative probability of that COA as it changes
over time. Remember the red COA is the "none of the above"
hypothesis.
The first thing that happens is that COA FIVE and COA FOUR
probabilities increase because of the failure to detect the first
two units. Then the one of the units is detected at H+1:15, driving
those COA probabilities down in favor of the other three. Next,
the NULL COA probability increases because of the failure to
detect another of the units, indicating that we may not know
what is going on. Finally, that unit is detected at H+2:15,
totally eliminating all COAs except the two routes through MOUNTAIN
PASS NORTH.
While we have only shown a sample here, we were successful
in creating a complete scenario for a recent project for the
U.S. Army Space and Missile Defense Command.
Contact
Dr. C. Allen Butler
for further information.
|
