You are on page 1of 8

COGENT: Cognitive Agent to Amplify Human Perception

and Cognition
Subrata Das and Dan Grecu
Charles River Analytics, Inc.
725 Concord Ave.
Cambridge, MA 02138, USA
+1 (617) 491-3474
{sdas, dgrecu} @cra.com

This is particularly the case in critical domains such as ground-


ABSTRACT based spacecraft operations, where mission operators receive
streams of information from a large number of sources: telemetry
In this paper, we demonstrate the feasibility of a cognitive agent,
data from the satellite, environmental data, information from other
COGENT, to amplify human perceptual and cognitive abilities in
control sites, or data stored in databases.
complex high-value operational environments such as are found
in spacecraft ground-based telemetry monitoring systems. The The amount and variety of incoming information creates the need
architecture of COGENT is based on Rasmussen's integrated of aiding the user in assimilating it, and of supporting him/her to
theory of human information processing, and supports decision- integrate the information in various types of task-specific
making at multiple levels of abstraction and complexity. The processing. While there are many tools available that separately
decision-aiding process ranges from situation feature extraction address information presentation or decision-support, there are
and summarization to support perceptual processing, through few approaches that set out to provide the user with integrated
situation assessment and alert generation to support enhanced support, at different levels of information abstraction and for a
situation awareness, to situation-specific response combination of tasks. This goal represents a particular challenge
recommendation to support complex decision-making. COGENT given that there is a tight corroboration between the information
displays information in a format that is easily understood, and contents, the format in which it the information is presented to the
matches the user's own mental model of the situation, thus user, and the tasks it can be used for.
requiring minimal amounts of translation and cognitive processing In this paper, we describe an agent-architecture developed to
for assimilation. The robust performance of COGENT is assured support users by amplifying perception and cognition in critical
through the use of complementary AI techniques such as operational environments. The goal of the agent development is to
probabilistic belief networks and argumentation. We developed provide the user with information that is timely, relevant, and easy
and demonstrated a prototype for COGENT, and tested it on real to assimilate. We first describe the background in which the agent
telemetry data sets. We are specifically addressing our effort for is intended to operate, and how users relate to this context. We
enhancing operator situation awareness and decision-aiding then introduce the COGENT architecture designed to facilitate
capability at ground stations. and enhance the user's interaction with the operational
environment. The description will further focus on the decision-
Keywords
aiding component within the agent, and will be followed by
Cognitive agent, decision aiding, event generation, situation implementation details, conclusions and future work directions.
assessment, information filtering, visualization.
2. BACKGROUND AND APPROACH
1. INTRODUCTION
The motivation for developing an agent that amplifies human
Modem decision-making in real-time dynamic environments is perception and cognition lies in the specifics of operator tasks,
increasingly becoming an information management task. The which require competent response through a range of behaviors
decision-maker must integrate domain knowledge gained from spanning from the reactive to the highly deliberative end of the
previous experience with the large quantities of incoming real- spectrum. Recent research indicates that expert decision-makers
time data, reflecting the current situation. quickly assimilate and integrate large quantities of information
and generate appropriate decisions if the information is presented
Permission to make digital or hard copies of all or part of this work for in a format that is appropriate for the level at which the
personal or classroom use is granted without fee provided that copies information will be processed (recognition-primed decision-
are not made or distributedfor profit or commercialadvantage and that
making) [ 1,6]. Humans are generally not good at handling large
copies bear this notice and the lull citationon the first page. To copy
otherwise, to republish,to post on serve~, or to redistribute to lists, amounts of low-level details, at keeping track of uncertain
requires prior specific permissionand/or a fee. information, at propagating it through mental simulations, and at
Agents 2000 Barcelona Spain maintaining unbiased perspectives in high-value outcome
Copyright ACM 2000 1-58113-230-1/00/6...$5.00 situations [ 15].

443
The control of spacecraft through ground operation centers is one problems. Examples of this type of processing include reasoning
of the domains where operators face these types of challenges. about unfamiliar faults in the spacecraft, and other similar tasks.
Satellites send streams of telemetry data to the ground center
describing the status of the onboard systems. The format of the
data consists of the parameter name or mnemonic (representing a
specific reading or onboard parameter setting), the parameter
value, the start and the end time for the associated values, and an
indication of whether the parameter's value is within normal
operating limits. Typically, such data streams can include Increasing
hundreds of messages within minutes. Some messages are critical lcompl
evelofexity
on their own, especially if they report the malfunctioning of a m
system. Other messages provide information through their informatinogn
processi
cumulative effect, indicating an evolution process. And there also / (e2RUxl
bl,bBl~~'kt.ai !edo°Pcuf~2~abltih~£or
are messages that have meaning only when used in conjunction
with additional messages indicating together the occurrence of a (e.g.,sensorimotor behavior)
high-level situation. The operator is required to identify and
recognize the relevant information in each of these classes, and to
INPUT OUTPUT
respond with appropriate decisions. ENVIRONMENT [
In the context of the tasks required from a ground control
Figure 1: Rasmussen's hierarchy of human cognitive
operator, providing support through an external decision-aiding
processing
system implies:
• Filtering out mission-irrelevant information, assessing By dividing skilled behavior into categories based on the degree
of automaticity, complexity, and level of cognitive processing, the
situations, and actively presenting mission-relevant
Rasmussen framework supports systematic analysis of the
information in direct correlation with the points reached in
information and knowledge required at each stage of processing.
the decision-making process;
In the context of COGENT, this framework provides guidelines
• Providing information to the operator that is difficult or time- for the design of the decision-aiding module, specifically, for
consuming to derive due to the natural limitations of human defining functionality to support the operator in three major tasks:
cognitive and perceptual processing; and data preparation, situation assessment, and response
• Presenting the derived information in a format that is easily recommendation. These decision-aiding functions and their
understood and that matches the operators own mental implementation in COGENT are discussed in the next sections.
model of the situation, thus requiring minimal amounts of
translation and cognitive processing for assimilation. 3. COGENT ARCHITECTURE
To address these objectives in an integrated fashion we have The underlying philosophy motivating our approach is to provide
developed an innovative agent architecture based on an integrated decision-support and visualization for complex high-value
theory of human information processing [14], which supports operations (e.g., ground station operation, spacecraft operations,
decision-making at multiple-levels of abstraction and complexity. flight operations, etc.) in a manner that complements existing
The architecture supports decision-aiding ranging from simple human perceptual and cognitive abilities in tasks that humans find
perceptual enhancement of salient situation features to support difficult. These tasks include the management of uncertain
perceptual processing, through event detection and situation information, keeping track of large numbers of details, detecting
assessment to support enhanced situation awareness, to situation- conflicts among currently active alternatives, preventing biased
specific response recommendation to support complex decision- use of existing information, and assisting in contingency planning
making and option selection. and action/reaction assessment, all of which have been identified
Rasmussen's three-tier model of human information processing, to be characteristic of expert decision-makers [3].
shown in figure 1, is well suited for guiding both the definition of Figure 2 illustrates the overall architecture of COGENT with its
the mission operator's information requirements and the task and two major components: the decision-aiding module and the
user profile [13,14]. The arch of the figure represents the flow of visualization interface module. The decision-aiding module maps
information through the human decision-maker. The left side of the hierarchical Rasmussen cognitive processing model, and is at
the arch corresponds to stimulus processing, the right side the core of COGENT providing three key functional components:
corresponds to motor processing. Processing is divided into three 1) The data preparation component receives the raw telemetry
broad categories, corresponding to activities at three different data, extracts relevant situation features, and generates events
levels of complexity. At the lowest level is skill-based reflecting the on-board changes. 2) The situation assessment
sensorimotor behavior, such as perceptual feature extraction and component generates evaluations based on the events produced
hand-eye coordination. This level represents the most automated, from the telemetry data, and generates alerts whenever the results
largely unconscious level of skilled performance. At the next level of the evaluation indicate critical states. 3) The response
is rule-based behavior, exemplified by simple procedural skills for recommendation component ranks possible operator decisions in
well-practiced, simple tasks such as detection of significant the current situation based on the associated context and on prior
events. Knowledge-based behavior represents the most complex knowledge. The three components of the decision-aiding module
cognitive processing used to solve difficult and unfamiliar are software components configured through a predefined

444
application programming interface. The modular structure task and situation characteristics. COGENT is designed to display
provides the possibility to plug in modules allowing COGENT to information relevant to the current situation, thus supporting the
interact with other instances of COGENT agents, with external user to focus attention on key attributes and relations.
information sources such as ground-stations, spacecraft and
satellites, databases and remote sensing equipment. 4. DECISION-AIDING IN COGENT
Our goal within this paper is to illustrate how COGENT integrates
several techniques to successfully support alternating reasoning
types of the human operator. We focus in this section on the
particular types of support that are being provided through
COGENT, and on how they all fit together in the overall task of
mission control.
Situation Fe~ures Alerts Recommendations
4.1 Data preparation
Spacecraft situation assessment and status visualization require an
appropriate processing of raw telemetry to extract situational
Decision features as events, to filter out irrelevant information, and to
Aiding summarize data to prevent information overloading (figure 3). The
event generation step maps the incoming data stream into a pre-
Module
defined set of domain relevant events. The information filtering
step then filters events according to the operator profile. The
summarization step summarizes telemetry events without any
significant loss of user perception about the status of the satellite.
Users can be presented with filtered events as well as their
summary.
To Situation Assessment
and Visualization Modules
A
Summarizeddata& eventa T

..... N ..........
I Fh'~t¢ldata & events
Figure 2: COGENT architecture

The decision-aiding module uses a combination of techniques to


ensure robust aiding with respect to the challenges raised in each
operational reasoning phase. At the lowest level, probabilistic
rule-based techniques are used for event generation and
information filtering in situations where incoming data is
!
uncertain, and/or conflicting. The decision-aiding module
maintains a database, which serves as a repository of the Payload I Status& health
data I data
accumulated filtered events from raw telemetry data, and which
stores user supplied profiles for information filtering and alarm Figure 3: Data preparation phases and their relation to
generation. At the next level, Bayesian belief networks implement situation assessment and information visualization
situation assessment under incomplete situation descriptions.
Finally, at the top level, the operator is supported in selecting
between several possible decision alternatives through a response 4.1.1 Event generation
recommendation mechanism based on argumentation techniques. The event generation step correlates the data from onboard
The visualization interface module displays the information sensors to specific pre-determined situation features, and
produced by the decision-aiding module in a format matching the generates low level situation events. For example, the temperature
typical representation for the task currently executed by the readings from a payload are fused with spacecraft attitude system
operator, and the operator's cognitive and decision-making style data to generate the 'change in temperature' event for the payload
and training. The information provided by the visualization when it is not pointing directly to the sun, and therefore a
interface is derived through the processing stages within the temperature rise is not expected. The event generation step is
decision-aiding module including telemetry summarization, further justified by the fact that continuous sensor readings must
argumentation for or against actions, and causal explanation for be quantized for the belief net (BN) processing that follows in the
situations that have been assessed. The agent's local database situation assessment phase.
stores task and user profiles that define the information display Events are defined as occurrences (e.g., actions or changes of
format matching typical decision-making patterns and specific state) in the domain that may affect or impinge on domain objects.

445
For example in engineering status and health data, events may telemetry values. As with the continuous events, the attributes for
include changes of values for discrete data types, large or small the discrete events are parameter Value, Start Time, and End
deviations in values for continuous data types, or out of limit Time.
values for all data types. Events have target objects and attributes
The event model that we have adopted implemented a flexible and
or parameters (e.g., for an out of limit event, the target would be
intuitive model which provided COGENT the necessary means to
the associated satellite component registering the data reading and
perform data filtering, summarization, and prioritization.
possible event attributes would be the time of the reading and
value of the reading). Events are discrete entities that indicate the
level of normality of the target's behavior. We used an event 4.1.2 Information f i l t e r i n g
model having event types based on individual telemetry parameter In the context of engineering data management, information
values. filtering is the mechanism through which the user/ground team
COGENT's event generation module operating on the available can subscribe profiles that are continuously evaluated, and allows
telemetry data uses the event model outlined in figure 4. As COGENT to send filtered events according to the profile. A
shown, we have two sets of event types (denoted by rectangles) profile is a set of rules of the form:
defined based on the target objects (denoted by ovals). The target
objects within our model are continuous-valued parameters and IF Spacecraft-Mode(X) THEN Filter(Y)
discrete-valued parameters. Individual event types are derived The mode variable X can take values such as normal or fail-safe.
from the color schemes existing within the telemetry files (e.g., The filter variable Y can be any combination from the set of
Green, Red, Yellow, etc.). values into which spacecraft parameters are discretized according
to their degree of emergency: High Red, High Yellow, High
Continuous Events
Green, Green, Low Green, Low Yellow, and Low Red.
Consider a profile consisting of the following two rules:

IF Spacecraft-Mode(normal) THEN Filter( { Green, Yellow})


IF Spacecraft-Mode(fail-safe) THEN Filter(None)
The first rule states that if the spacecraft mode is normal then
/ / J ~'.~ Green and Yellow events should be filtered out, and only the Red
Value End Time events should be passed to the data summarization module. On
the other hand, the second rule states that everything passes
Discrete Events through the filter if the spacecraft is in fail-safe mode.

Value
/
O
Start Time End Time
4.1.3 Data summarization
Summarizing status and health data or events from a filtered data
or event stream aims to save bandwidth without any significant
loss of operator perception about the status of the satellite.
COGENT uses several classes of techniques for summarizing
events such as saliency, abstraction, integration, link analysis, and
statistical analysis. We exemplify two some of them here,
Figure 4: Event model used by COGENT indicating the way they support mission analysis at a semantic
level.
For continuous-valued parameters, seven event types are present
COGENT computes correlation coefficient matrices across
ranging from High Red to Low Red. Green values indicate
parameters, by determining the linear relationship between two
parameter levels that are within the nominal range, Yellow
parameters x and y as measured by the Pearson coefficient of
indicates tolerable out-of-range values, while Red stands for
linear correlation r [9]. This method is similar to a zone frequency
signal deviations-that correspond to an emergency level. High and
analysis except that the warning limits are statistical, such as lcr-
Low specify whether the qualification refers to the superior or
inferior border of the parameter range. Whereas the range for Red level, 2G-level and 3a-level, where ~ is the standard deviation of
and Yellow events are pre-set based on existing spacecraft limit the parameter values. A correlation between parameters is a data
settings, the ranges for the Green events are user-specified. For summarization form stipulating that if the agent sends information
example, the user can choose to evenly divide the Green region for x, it should also send information for y. For example, to
(33.33% each), or make the Green region take up 90% with Low summarize a High Red event for x, associated data for y should
Green and High Green taking the lower and upper 5% each of the also be included.
nominal green range, respectively. Attributes for each continuous A second summarization method used by COGENT is event
event, as shown in figure 4, include Value, the time the parameter clustering, as implemented through time variant, time invariant,
took that value (Start Time), and the final time the parameter had and delta methods. The time variant method puts consecutive
that value (End Time). events within the same zone (e.g., High Yellow) into one cluster.
Event types for discrete-valued parameters include Green and Each of the clusters is represented by three parameters: its size,
Red, corresponding directly to the information given in the the mean value, and the time-period over which the cluster
expands. The time invariant method clusters events according to

446
the zone. Each such cluster is represented by its size and the mean also to the response recommendation module, even if their
of all the values falling within the cluster. This clustering method criticality does not require alert generation.
is therefore a generalization of the zone frequency method. The
The high-level situation assessor is implemented using the
delta interval method does clustering around events that are
Bayesian Belief Network (BN) model [8,11,12]. Besides being
considered important or unusual. The delta interval method allows
computationally tractable, and besides capturing the cause-effect
one to analyze the immediate predecessors and successors of the
relationships that exist amongst the variables of the domain, belief
event being considered. The clustering summaries are critical in
network models offer the advantage of closely reflecting expert
providing the operator with parameter trend analyses, and in
knowledge. This typically makes them easier to construct (thus
determining situation criticality as transient or stabile depending
minimizing the knowledge engineering costs), and easier to
on the time spans covered by the clusters. Summaries will allow
modify.
to further extract symbolic information from the data that can be
used in prioritizing the information presentation process for the The COGENT situation assessment component uses belief
operator. networks to verify spacecraft subsystems status following operator
commands. When a command is sent to the spacecraft, the belief
As in the case of event generation and filtering, summarization is
networks determine the expected values of relevant status
driven by the profiles selected by the user as being appropriate for
parameters. The resulting state is then compared with the actual
the task currently in progress. Profiles can be changed in response
telemetry to detect anomalies.
to changes in the mission development or in the goals that are
pursued.

4.2 Situation assessment


The next level of intelligent operator support through COGENT is
intended to provide a situation analysis based on the semantic and
causal integration of the events generated on board of the
spacecraft. The purpose of this functionality within COGENT is
twofold. First, given that the agent acts as a gateway for the
information streamed towards the operator, it has a much better
opportunity than the human operator to capture the data that is
part of a causal model. Consequently, COGENT can continuously
update the model, and keep the user informed only on the high-
level aspects that are of immediate interest. Second, COGENT has
a much better potential to monitor values that result through the
assessment, to identify abnormalities, and to use the model to Figure 6: Belief network for spacecraft anomaly detection
identify potential causes. Figure 6 provides a belief network example for command
verification. Each variable in the network is a parameter
1• A High-Level representing either a subsystem parameter (e.g., P Y M s p e e d -
wheel speed, N S D S V , north-south downstream valve) or a
command (e.g., MDmaneuver- momentum dump maneuver,
ThrActiv - thruster activation). The situation illustrated in the
figure reflects a case where a high wheel speed causes a value
Event. High'!Tevel
].,.~l indication of PYMspeed in the High Red range. The action can be
Situations corrected through a momentum dump maneuver (MDmaneuver)
which will reduce the excessive speed. The momentum dump
maneuver in turn requires a thruster activation (Thr._Activ). Thus,
' ~ High-LevelSituationAssessor ]~ following along the direction of the red arrows in figure 6, the
observation of a high value of PYMspeed should automatically
Events recommend the thruster firing. If this command is sent to the
spacecraft then the expected state for the north-south downstream
Figure 5: Situation assessment component in COGENT valve (NSDSV) should be off. If the state of NSDSVis found to be
on in the incoming telemetry then an anomalous situation will
The diagram in figure 5 summarizes these two functions as they occur, and an alarm should be generated.
are implemented by COGENT. The input to the situation This situation is captured in figure 7, a snapshot from COGENT's
assessment component is represented by the low-level events execution. The red arrow on the left-hand side of the figure
generated within the data preparation module. The events may points to the expected value (off~green) computed by the belief
directly trigger alerts if they exceed pre-determined limits. Events net. The red arrow at the bottom of the figure points to the actual
are simultaneously input into the high-level situation assessor, telemetry reading (on~red). The anomaly is detected and therefore
which is also monitored by the alert generator. The alerts that are an alarm is generated (at the center of the figure).
generated this time refer to high-level situation evaluations. The
results of the situation assessment are provided to the user, and

447
Situation
- temperature increase in imaging payload
Possible responses
- slew payload
- cover payload
Example Arguments for and against candidates
IF imaging not urgent AND temperature is acceptable
THEN cover payload (sl)
IF temperature is too high
THEN slew payload (s2)
IF covering process takes exceeds payload exposure limit
THEN slew payload (s3)
IF payload will soon be out of the sun
THEN NOT slew payload (nl) AND NOT cover payload (nl)
where sl, s2, s3, and nl represent strengths of arguments
Figure 7: Expected value and actual telemetry comparison for/against the responses
followed by alarm generation
Final recommendations and explanation:
- slew payload
4.3 Response recommendation
The overall function of the response recommendation module is to Table 1: Example of Response Recommendation Process
recommend a suitable action based on the current situation, as
identified by the situation assessment module. The The evaluation of the arguments is performed through a process of
recommendation process implements Toulmin's model of aggregation. This process includes not only an assessment of the
argumentation [16]. The model has been successfully applied in number of arguments and the strength of each independent
various application domains including medicine [2,5] and argument, in other words is based on the collective force of
software verification [4]. The internal architecture of the response arguments for each hypothesis, but also on the plausibility of the
recommendation module is shown in Figure 8. evidence and justification for each line of argument. For example,
if the current situation provides two arguments for the response
"slew payload", then these two arguments will be aggregated to

l Recommendation

Aggregation ]
produce a combined strength, which is expected to be higher than
both s2 and s3. If we use the probabilistic dictionary [0, 1] to
represent strength, then the two arguments can be aggregated by
using a special case of Dempster's epistemie probability [7] which
is s2 + s3 - s2 * s3. On the other hand, if we use the dictionary

[
T Arguments
Argumentation ]
[-1, +1], then the arguments can be combined using formulae
from the certainty factor formalism.
For each response all arguments for and against it are combined in
~ Decision Options the above manner to accumulate total evidence for and against the
response. Evidence against each response discounts evidence for
I Decision Option Generator I the response. For example, if a particular situation provides
arguments for and against the response "cover payload" with the
TEvents and Situation Types evidence being respectively s l and n l , then the aggregated
Figure 8: Internal Architecture of the Response evidence is sl*(1 - nl)/(1 - s l * n l ) . If such an aggregation process
Recommendation Module associates more weight to the response "slew payload" than to any
The first step of the response recommendation process identifies other response, then the response will be considered as a best
all relevant candidate response options in the current situation. option in the current situation, and will be presented as such to the
The second step then generates and evaluates arguments for and operator.
against these options. The final step chooses the best response A key aspect of COGENT's decision-aiding process is the ability
option by evaluating these arguments. to explain the user the system output by following the steps of
The knowledge used to construct and evaluate arguments is arguments. This is particularly critical when actual decision
represented by rules. Examples of rules for the current scenario recommendations are being made. COGENT provides
are shown in Table 1. The values of the strengths of arguments for justifications for its decision recommendations in the form of
( s l , s2, s3) and against ( n l ) a response are obtained through a high-level inferencing diagrams illustrating the data sources and
knowledge elicitation process with expert decision-makers. knowledge used to derive any conclusions or recommendations.

448
5. E V A L U A T I O N AND complex situation assessment and anomaly detection by modeling
the problem into a network structure that embeds key domain
IMPLEMENTATION concepts and relationships used by ground operators and
COGENT was tested on a data set consisting of eight different spacecraft analysts. We have used our in-house belief network
telemetry passes. Of these eight passes, seven encompassed engine to implement the situation assessment module of
normal operation while one pass included an orbital maneuver. COGENT.
Two event profiles (specifying how parameter values are mapped The response recommendation module is implemented on a
to events) and three filtering profiles (determines which event standalone Prolog platform. The implementation currently
types are allowed to pass through) were selected. In general, each incorporates two dictionaries: the qualitative dictionary {% ++}
pass generated on the order of several thousands of events. and the probabilistic dictionary [0, 1]. We plan to enhance the
Filtering reduced the event streams to as low as thirty or as high number of dictionaries and integrate the current standalone
as four hundred events based on the event/filtering profiles that argumentation module with the COGENT prototype.
were active. Summarization results concentrated on a comparison
of event summaries between the non-normal pass and a normal 6. C O N C L U S I O N S AND F U T U R E
pass. Statistical and clustering summarization techniques clearly DIRECTIONS
differentiated between the two passes by showing the higher level
variability and fluctuation amongst event types for individual In this paper, we have presented a cognitive agent, COGENT, to
parameters for the non-normal data set. amplify human perception and cognition. The COGENT agent is
motivated by cognitive theory to provide decision-aiding at
The situation assessment process compared the actual spacecraft
multiple levels of information processing. COGENT relies on
state with the expected state and generated alarms in case of
complementary AI techniques to support processing at three
anomalies. The visualization module mapped the filtered
levels: data preparation via data fusion, information filtering and
summaries into high, normal, and low priority queues, with
summarization techniques, situation assessment via belief nets,
summaries relating to important attitude control system
and response recommendation via argumentation. We have
parameters being automatically included in the high priority
demonstrated the use of COGENT in the context of a ground-
visualization queue. The events for situations that were evaluated
based control for communication satellite problem using telemetry
as anomalous were further forwarded to the response
data sets. The scope of this paper is the decision-aiding aspect of
recommendation module.
COGENT. The other characteristics of COGENT including,
The visualization module provides the operator with information visualization, communication, and learning, will be detailed
resulting simultaneously from all three levels of decision support. elsewhere.
Figure 7 illustrates a situation where the operator is presented
We are currently exploring further development directions for the
with time variant summaries for the NSDSV (north-south
decision-aiding module towards an enhanced functionality and
downstream valve activation) parameter, with the value for the
potential extension to other domains.
corresponding belief network node, and with the alarm that is
generated as a result of comparing the expected value with the One first goal is to transition the level 1 data fusion manager
actual telemetry reading. The visualization module ranks the towards fuzzy logic (FL) technology. Fuzzy logic [17] provides
output of the three decision-aiding modules based on the current the necessary human-like reasoning and handling of uncertainty to
situation, and using knowledge about the information expectations manage and coordinate date fusion components and sensor and
of the user in that particular situation. For example, COGENT information assets. While some event detection can be
maps filtered and summarized data into one of three queues: high implemented via simple discrete Boolean logic, many onboard
priority, normal priority, and low priority. The mapping or events will require a more robust and flexible means of definition.
visualization profile is specified via a user-set list of high priority This can be achieved by the use ofa FL-based event detector.
parameters, an acceptable number of limit exceedances, and On the situation assessment side, one major problem we are
allowable thresholds for outlier values. The visualization of confronted with is that the type of belief network currently used in
summarized data involved the display of clustering, parameter COGENT does not model temporal information explicitly. We
data distribution, temporal trends, etc. Visualization of belief currently plan to extend the belief net framework to incorporate a
network inferencing involved the display of the current states of means for representing time-varying or dynamic spacecraft status
the variables in the network concerned. The system visualized the information. The essential idea underlying temporal or dynamic
state of a node change, upon the propagation of a control belief networks (DBN) [10], illustrated in figure 9, is that
command or the propagation of a telemetry data item. knowledge of the previous state of the belief net renders
The data preparation module is implemented as a message parser information about all preceding states irrelevant. In other words,
of telemetry messages, and uses an object based representations to the current state t, of the belief net is conditionally independent of
represent the encoded state changes. Simple subsytem models are all states prior to state t-1 given the state at t-1. (Note that the
used to encode the evolution dynamics as a result of commands nodes of the DBN of figure 9 stand for collections of nodes
and of the generated events. The event filtering component is rule- having the same timestamp.) The evidence propagation
based, using primarily Boolean filtering criteria. The mechanism in a DBN requires no modifications from that of static
summarization component relies on a library of statistical and belief nets.
clustering methods.
The situation assessment module is developed based on a library
of Bayesian belief network algorithms. Belief networks support

449
State EvolutionModel [4] Forder, J., C. Higgins, et al.. SAM - A tool to support the
construction, review and evolution of safety arguments.
Proceedings of the Safety-critical Systems Symposium,
London, Springer-Verlag, 1993.
[5] Fox, J. and P. Krause. Qualitative Frameworks for Decision
Support: Lessons from Medicine. The Knowledge
Engineering Review 7(1): 13-33, 1992.

SensorModel [6] Klein, G. A. Recognition-Primed Decisions. Advances in


Man-Machine Systems Research. W. B. Rouse. Greenwich,
Figure 9: Temporal (dynamic) Belief Net CT, JAI, 47-92, 1989.

We have implemented an argumentation-based decision making [7] Krause, P. and D. Clark. Representing Uncertain Knowledge,
process as detailed above and have tested it with the example in Intellect, Oxford, England, 1993.
Table I. The example itself was constructed in consultation with [8] Lauritzen, S. L. and D. J. Spiegelhalter. Local Computation
our Subject Matter Experts (SME). During knowledge elicitation with Probabilistics in Graphical Structures and Their
sessions the evidence attached by the SME to an argument was Applications to Expert Systems. Journal of the Royal
not always expressed as a probability value; rather the SME has Statistical Society B 50(2), 1988.
frequently used terms like "high", "low", "possible", "likely", etc.
We consider this issue to be an important candidate extension for [9] Matlack, W. Statistics for Public Managers. F.E.Peacock,
the argumentation component. Itasca, IL, 1993.
A key aspect of our decision-aiding process is the ability to [ 10] Nicholson, A. E. and Brady, J. M. Dynamic belief networks
explain the system output to the user by following the steps of the for discrete monitoring. IEEE Transactions on Systems, Man,
arguments. We are currently exploring the further integration of and Cybernetics 24, 1593-1610, 1994.
this requirement with the visualization module to develop
[ll]Pearl, J. On Evidence Reasoning in a Hierarchy of
visualization formats that depict both the outcome and the
Hypotheses. Artificial Intelligence 28, 9-15, 1986.
structure .of the arguments, and to allow the operator maximal
insight into the decision-aiding process. [12]Pearl, J. Probabilistic Reasoning in Intelligent Systems:
Networks of Plausible Inference. Morgan Kaufinann, San
7. ACKNOWLEDGMENTS Mateo, CA, 1988.

This work was performed under contract NAS2-99018 with the [13]Rasmussen, J. Skills, Rules and Knowledge: Signals, Signs
NASA Ames Research Center in Moffett Field, California. The and Symbolism, and Other Distinctions in Human
authors thank Bill Clancey from the Human-Centered Computing Performance Models. IEEE Transactions on Systems, Man,
Group at the NASA Ames Research Center, Greg Zacharias, Paul and Cybernetics 12, 257-266, 1983.
Gonsalves and Rod Rinkus of Charles River Analytics, and Eva [14] Rasmussen, J. Information Processing and Human Machine
Hudlicka of Psychometrix for their invaluable contribution and Interaction: An Approach to Cognitive Engineering. New
aid during the project. York, North Holland, 1986.

8. REFERENCES [15]Tolcott, M. A., F. F. Marvin, et al. Expert Decision-making


in Evolving Situations. IEEE Transactions on Systems, Man
& Cybernetics 19(3), 606-615, 1989.
[1] Brezovic, C. P., G. A. Klein, et al. Decision-Making in
Armored Platoon Command, Klein Associates, 1987. [16] Toulmin, S. The Uses of Argument. Cambridge University
Press, London, England, 1958.
[2] Das, S. K., J. Fox, et al. Decision Making and Plan
Management By Autonomous Agents: Theory, [17]Zadeh, L. A. Fuzzy Sets. Information and Control 8, 338-
Implementation, and Applications. First International 353, 1965.
Conference on Autonomous Agents, ACM Press, 1997.
[3] Fallesen, J. J. and R. R. Michel. Observation on Command
and Staff Performance During CGSC Warrior '91, US ARI
for Behavioral and Social Sciences, 1991.

450

You might also like