A Comparison of Deterministic vs Stochastic Simulation Models for Assessing

Adaptive Information Management Techniques over

Disadvantaged Tactical Communication Networks

Dr. Allan Gibb and Mr. Jean-Claude St-Jacques *

Defence R&D Canada - Valcartier
2459 Pie-XI Boulevard North
Val-Bélair, QC, CANADA G3J 1X5
1 (418) 844 4000 ext 4683, ext 4376
allan.gibb, [email protected]

and

Dr. Gerard Nourry and Mr. Tim Johnson

IP Unwired Inc.
Bldg 94, 3701 Carling Avenue Suite 228
Station H, Box 11490
Ottawa, ON, CANADA K2H 8S2
1 (613) 829 1310 ext 228, ext 233
gerardnourry, [email protected]

Abstract

A test bed is being designed and implemented at DRDC Valcartier1 to study the effect of
information management techniques, applied at the level of the application database in each
participating node of a simulated tactical radio network, on the quality and timeliness of
information distributed across the network. This paper describes key concepts and architecture for
a test bed that was conceived to support these information management studies in a realistic
simulated battlefield environment. Two basic simulation models were considered for the test bed.
In the first, the battlefield and communications components of the simulation model are fully
deterministic. The battlefield component is implemented as an event-rich script consisting of an
invariant sequence of battlefield events. The communications component is based on a single
throughput-delay curve or family of such curves. In the second simulation model, the battlefield
component remains deterministic while the communications component introduces a stochastic
element to produce a more realistic characterization of the communication channel(s). The
communication model is based on a Markov process characterization of the communication
link(s). This gain in accuracy is realized at the expense of introducing a probabilistic element in
the model. The advantages and disadvantages of each approach are discussed in the context of
the information management studies they must support.
*
This work is sponsored by the High Capacity Tactical Communications Network Technology Demonstration
(HCTCN TD) project of Defence Research and Development Canada.

1
formerly known as DREV (Defence Research Establishment Valcartier); name change effective 1 April 2002.
1. Introduction

On the tactical battlefield, the means of communication between vehicles or dismounted soldiers is
generally a line-of-sight radio operating in the VHF or UHF band. The advent of tactical digital
command and control systems has generated a requirement to distribute significant quantities of
digital data over these radio systems. Initial experience with passing tactical data in digital form
over Army radios has highlighted the inadequate data capacity of the wireless channels. Armies
are learning that, on an information-hungry battlefield, demand for information in digital form will,
at least for the foreseeable future, exceed channel capacity.

One of the major reasons for introducing digital command and control systems into the tactical
domain is that they promise increased battlefield awareness through a more systematic and
automated distribution of relevant data than is possible with a voice-based communication system.
To deliver on this promise, however, the communication infrastructure must be capable of
distributing digital data among participating mobile command and control nodes with high fidelity
and with a timeliness appropriate to the operation (i.e. it must meet the commander’s critical
information requirements). To be useful, critical information must be passed quickly enough to
permit the friendly commander, and the staff that he commands, to stay and act within the
decision cycle of the enemy commander.

In the tactical domain, sharing of information on an ‘all-informed’ basis is highly desirable, if not
essential, to avoid a single point of failure and ensure continuity of operations. Under this
exchange model, neighbor nodes attempt to maintain exact copies of each other’s database via
asynchronous data replication. When the communication channels are characterized by low and
variable throughput and unreliable connectivity, maintaining exact copies of each other’s database
(‘full synchronization’ of database content) via asynchronous data replication can be impossible.
Due to the imperfect communication channels, it is likely that, over time, the databases will drift
out of synchronization and the users will be unaware that this is happening. Users will believe
that they are sharing the same situation picture, when, in fact, they are not. Such a characteristic,
if detected, can undermine confidence in the system. Undetected, this characteristic may have
deadly consequences.

In such a communications environment, it is important to build into the command and control
system information management strategies that can adapt to changing battlefield and network
conditions without user intervention so as to optimize the flow of priority and high-value
information between nodes regardless of the state of the communication channel.

At Defence R&D Canada - Valcartier, a test bed is being implemented to study the impact of
information management techniques on the quality and timeliness of information distributed across
the tactical battlefield [Gibb and St-Jacques, 1999]. From a quality of information perspective,
the consistency of the information stored and used at various levels of command is a major area of
concern. The test bed will support the development and execution of Concept Development
Experiments (CDEs) aimed at improving information dissemination and information quality on the
battlefield. Key components of this test bed include a battlefield model and a tactical radio
communication model.
Two basic simulation models were considered for the test bed. The first is characterized by
battlefield and communications components that are fully deterministic. In the second, the
battlefield component remains deterministic while the communications component introduces a
stochastic element to produce a more realistic characterization of the communication channel(s).
This paper discusses the advantages and disadvantages of each approach in the context of the
information management studies each must support.

2. Terminology

The terminology definitions below are taken from [Banks and Carson, 1984].

Discrete-event systems simulation is the modelling of systems in which the state variable changes
only at a discrete set of points in time. The simulation models are analyzed by numerical methods.
A simulation model is a particular type of mathematical model of a system. Models can be
classified as static or dynamic, deterministic or stochastic, and discrete or continuous.

A static simulation model, sometimes called a Monte Carlo simulation, represents a system at a
particular point in time. Dynamic simulation models represent systems as they change over time.

Simulation models that contain no random variables are classified as deterministic. Deterministic
models have a known set of inputs which will result in an unique set of outputs. A stochastic
simulation model has one or more random variables as inputs. Random inputs lead to random
outputs. Since outputs are random, they can be considered only as estimates of the true
characteristics of a model. In a stochastic simulation, the output measures must be treated as
statistical estimates of the true characteristics of the system.

A discrete model is one in which the state variable(s) change only at a discrete set of points in
time. A continuous model is one in which the state variable(s) change continuously over time.

This paper describes a discrete-event system simulation in which the simulation model is dynamic
and discrete, with both deterministic and stochastic components.

3. Information Management Techniques

The information management techniques being studied are techniques that will exploit and act on
data in the local database within a tactical node. The techniques can take two basic forms - those
which prepare information in the most efficient possible form for transmission, and those which
limit what is transmitted and/or when it is transmitted. As simple examples, a data compression
algorithm would fall into the first category. A rule which inhibits the broadcast of an ‘own
position’ report from a vehicle whose position has not changed would fall into the second
category. For the most part, the techniques will be implemented as expert rules tied to individual
data elements in the database using the mechanisms of triggers and stored procedures. The
purpose of the expert rules is to build into the tactical nodes information exchange control
mechanisms which will adapt themselves without user intervention to changing operational
context and/or to changing conditions on the communications network.

In principle, it is possible to analyze each information management technique individually and to

project its impact on the battlefield for certain limiting cases. However, the information
management techniques are adaptive in nature and intended to be applied together. The
combined effect under realistic operational conditions is extremely difficult to assess unless those
conditions are simulated with some degree of fidelity. What is required is an infrastructure which
permits a) realistic tactical scenarios to be played out, b) behavior of the communications channels
to be simulated with fidelity, c) information management techniques to be applied throughout the
scenario run, and d) data to be recorded during the scenario run which will allow the impact of the
techniques to be assessed. The test bed will provide that infrastructure.

4. Test bed Concept

4.1 Architecture

The test bed will consist of a set of simulated tactical nodes (maximum thirty-two, four shown)
linked together through a central server node [Figure 1]. The latter node will contain a software
module that simulates relevant characteristics of the wireless communication system. Scripted
tactical scenarios, spanning a spectrum of tactical Army operations, will be played out which
require information in the form of data transmissions to be passed between simulated nodes
through the communications system simulator. The simulated nodes will have just the
functionality necessary to mimic relevant functionality of true tactical command and control nodes
(e.g. a geo-referenced map display with moveable icons). The scripted scenarios will be capable
of automated execution without men in the loop through use of scripts, or manual execution with
men in the loop stationed at the consoles of the simulated tactical nodes.

Simulated command
and control nodes

Tactical Node wireline link Tactical Node

Central server node

(Radio communication
simulator)

Tactical Node Tactical Node

Figure 1. Test Bed with Radio Communications Simulator

The test bed will permit a second mode of operation (Figure 2) in which a tactical radio is
interfaced to each node and the data communications related to scenario events are broadcast
over the radio system. In this mode, the wireline links between the nodes and the central server
are maintained to ensure synchronization of scenario execution across nodes and to enable
inspections of database content and data logging during pauses in scenario execution.

Simulated command
and control nodes

Tactical Node Tactical Node

wireline link

Central node

Tactical Node Tactical Node

combat net radio

Figure 2. Test Bed with Tactical Radios

4.2 Information Exchange Model

In the tactical domain, sharing of information on an ‘all-informed’ basis is highly desirable, if not
essential, to avoid a single point of failure and to ensure continuity of operations. If neighbor
nodes have exact copies of each other’s database, a node can assume a neighbor node’s role,
when necessary, without requiring a massive one-time data transfer. As well, a node disconnected
from the network can recover lost data from any neighbor node upon reconnection.

Information exchange on the tactical battlefield based on database updates, as described by

Chamberlain [Chamberlain, 1996], offers the potential for a substantial improvement in the
utilisation of available bandwidth and strong support for nodal autonomy. With this approach, the
database schema supports a situational model of the battlefield. Each local node maintains its own
situation model in its database. The primary role of tactical data communications is to update each
other’s database rather than to exchange messages. In a bandwidth-limited environment,
database-to-database exchange through replication of database transactions is generally more
bandwidth efficient than the exchange of structured military messages. Database replication
exchanges only new or modified data while military messages, which must be semantically
complete, tend to exchange both modified and unmodified data.

The test bed will implement an ‘all-informed’ exchange model involving asynchronous replication
of database updates.
4.3 Effect of Communications Channel

In an ideal communications environment, characterized by perfect connectivity and zero latency,

asynchronous replication will occur with almost zero latency and database content will
synchronize across nodes with almost zero latency – a true ‘all-informed’ model.

In a non-ideal communications environment, connectivity is imperfect, bandwidth is finite, and

latency is non-negligible. Use of asynchronous replication means that a portion of the database
content at any given time will not be synchronized (i.e. values for certain data elements will not be
consistent across all network nodes), An efficient replication mechanism, given reasonable
connectivity and adequate bandwidth, can ensure that synchronization will occur with low latency
and that the portion of unsynchronized data will be kept small and relatively constant. Users will
not have exact copies of each other’s database, but will have almost-exact copies.

When communication channels are characterized by low and variable throughput and unreliable
connectivity (the subject of this paper), maintaining exact or almost-exact copies of each other’s
database via asynchronous data replication is not realistic. There is a risk that the databases will
drift increasingly out of synchronization and the users will be unaware that this is happening.
Users will believe that they are sharing the same situation picture, when, in fact, they are not.

4.4 Measurements

The purpose of the automated information management techniques is to ensure that, under the
conditions just mentioned, the tactical nodes will manage information exchange intelligently so as
to (1) minimize the quantity of data transmitted, and (2) ensure that data values for information of
high operational importance remain fully synchronized across the nodal databases.

Quantitative assessment of the impact of these techniques requires that the test bed provide a
controlled environment for the execution of experiments, i.e. that experimental conditions and
results be highly reproducible. Measurements made using the test bed infrastructure must provide
information about database states during scenario execution for the following three cases:

(1) Perfect communication channels;

(2) Communication channels characterized by low and variable throughput and unreliable
connectivity, with no intelligent information management techniques applied;
(3) Case (2) but with intelligent information management techniques applied.

Differences in data values for cases (1) and (2) provide a quantitative measure of the effect of the
imperfect communication channels on the de-synchronization of database content.

Differences in data values for cases (2) and (3) reflect the success of the information management
techniques in counteracting the negative effects of the communication channels.

Both sets of difference measurements are important. It is necessary to have a quantitative

understanding of the effect of the communications channel on database de-synchronization for
different communications traffic conditions and states of network connectivity before the positive
impact of the information management techniques can be reliably evaluated.

5. Deterministic vs Probabilistic Battlefield and Communications Models

Clearly, it is not possible to make meaningful measurements for all three of the above cases during
a single scenario execution. To measure the differences referred to above, it must be possible to
repeat the scenario several times. Two approaches are considered viable.

In the first approach, the battlefield model described in the scenario is fully deterministic (i.e.
capable of generating the identical sequence of battlefield events each time), as is the
communications model. With this approach, the scenario is repeated with perfect fidelity three
times, once for each of the above cases. It is then possible to attribute the difference in observed
results uniquely to the effect of the applied information management technique, because all other
experimental parameters are identical from case (2) to case (3).

In the second approach, one or both of the battlefield component and communications component
of the simulation model are stochastic in nature. Each time the simulation is run it will have a
slightly different outcome, even if the inputs are the same. This characteristic makes it impossible
to correlate differences in observed results uniquely with the effect of the applied information
management techniques. To observe statistically-meaningful differences, it would be necessary to
repeat each of cases (1) through (3) many times and to perform a statistical averaging so that
randomly occurring effects are averaged out. Such an approach significantly increases the amount
of effort required, both to run the experiments and to analyze the results.

For the battlefield component of the simulation model, the primary advantage of a stochastic
model is that such a model more accurately characterizes the unpredictable nature of the tactical
battlefield. It also helps to ensure that observed results are not unduly biased by the particular
choice of events comprising the scenario. However, the risk of biased results from use of a
deterministic battlefield model can be reduced by ensuring that the model contains a rich mix of
events judiciously selected to minimize such bias. Also, where bias is suspected, it may be
possible to test for bias by modifying a portion of the scenario or by introducing stochastic
behavior for certain variables, then re-running the experiment. A fully deterministic battlefield
model was selected for the test bed experiments because the advantages of full control over the
sequence of scenario events and the repeatability of those events far outweighed any
disadvantages.

For the communications component of the simulation model, on the other hand, one is trying to
characterize a real physical system whose behavior has an intrinsically random component. In this
case, the risk of obtaining biased results through use of a simulation model which does not
accurately characterize the behavior of the communication channel is a more serious issue.
Options for modelling the communications channel behavior, and the advantages and
disadvantages of each option, are discussed in Section 7.
6. Deterministic Battlefield Models

The battlefield model consists of a list of events occurring in a specific sequence on the battlefield
in the context of a tactical scenario, e.g. a tank moving from position A to position B, or a radio
communication between vehicle X and vehicle Y. A deterministic battlefield model has the
property that the sequence of events remains invariant each time the scenario is executed.

There are two basic approaches that can be used to generate a deterministic battlefield model. In
the first approach, one or more military subject matter experts (SMEs) develop the broad outline
of a tactical scenario, and the SMEs then generate the detailed script of battlefield events and
event timings to populate the scenario. Each battlefield event and (where appropriate) event
timing is captured into a script file. A software tool then reads the script file and converts the
scripted events into a format that can be understood by the communications component or
information management component of the test bed. In the second approach, a software
application that allows simulation of battlefield activities is used to generate the series of events.
If the application can be interfaced directly to the test bed through an interface module, the
generated events can be used to stimulate directly the communications component and
information management component of the test bed. If the application cannot be interfaced
directly to the test bed, or the simulation is not directly useable because it has a stochastic
component, the events and event timings can be captured in a log file. It may then be possible to
interface the log file to the test bed through a software module and to use the log file as the
source of the deterministic battlefield model. If the exact event sequence generated by the
application is not what is desired, but the log content can be cast into human-readable form, a
subset of the logged events can be read by subject matter experts and (following the first
approach described above) used to generate a script file.

Two software applications that may be useful for generating the deterministic battlefield model
are JANUS and ModSAF (Modular Semi-Automated Forces). Both have stochastic components
and so cannot be used directly to stimulate the communications component and information
management component of the test bed. However, log files which they generate can be useful.

6.1 JANUS

JANUS is a warfighting simulation that challenges commanders to plan and interactively fight
battles against real world opponents. JANUS can see the opposing sides, but the players cannot.
Commanders plan and execute their battles on digitized maps anywhere in the world. The
warfighter's view is a two-dimensional map-like display that includes grid lines and graphic
control measures. To the weapon systems in the field, however, the terrain is three-dimensional,
with elevation and contour affecting maneuverability and lines of sight. Individual weapon systems
engage the enemy with the appropriate weapon or ammunition based upon range, target type, and
ammunition remaining. JANUS is an interactive, two-sided, closed, stochastic, ground combat
simulation. Data concerning discrete battlefield events and the position and status of weapon
systems are logged continuously throughout the scenario. These data can be used to replay the
scenario for post-mortem analysis. It is not clear at time of writing if scenario events recorded in
the JANUS database can be exported to another system in an useful format.
6.2 ModSAF2

ModSAF is an interactive, high resolution, entity level simulation that represents combined arms
tactical operations up to the battalion level. It is comprised of a set of software modules and
applications used to construct Advanced Distributed Simulation (ADS) and Computer Generated
Forces (CGF) applications that provide a credible representation of the battlefield, including
physical, behavioral and environmental models. ModSAF is based on two types of modeling,
deterministic and stochastic. Deterministic modeling in ModSAF is applied to the defined entities
behavior (e.g. attack the enemy if the distance to the target is less than a fixed value). Stochastic
modeling is applied to events such as detecting and firing on an enemy. Because of the stochastic
component, two runs with the same entities and the same initial values will result in different
simulation activities. The ModSAF data logger records the simulation packets of any protocol
family transmitted on the simulation network. This capability could permit the creation of a
scenario file that could generate an invariant battlefield model for all planned test bed runs.

7. Communications Models

The role of the test bed’s Radio Communication Simulator (RCS) is to approximate the
impairments that the battlefield communication environment would cause on the data/packet flow.
Information representing the battlefield communication environment must be fed into the
simulator in order that it can produce the associated impairments in the information flow.

The RCS consists of two main elements: the models used to simulate (1) the network
communication protocols and (2) the radio communication channel. Network communication
protocol models are built into most simulation packages and will not be discussed further here.
The radio communication channel model is emphasized in this paper. The channel model
characterizes the external (communication) environment in which a wireless system/network
operates, i.e. the communications channel between a pair of antennas. The channel accounts for
propagation effects such as path loss, rain absorption, multi-path, fading, diffraction, refraction,
and scattering, as well as general background noise. In the widest sense, the channel may also
account for sources of interference when these are treated in aggregate.

There are three broad categories of channel models: discrete, continuous and ray tracing models.
Primarily for reasons of computational efficiency, a discrete channel model approach has been
selected for the Radio Communication Simulator.

2
In the near future, the American Army will retire ModSAF and replace it by OneSAF. OneSAF is a next
generation Computer Generated Forces (CGF) that can represent a full range of operations, systems, and control
processes from the individual combatant and platform level to battalion level. Unit behaviors will be modeled to
the battalion level for selected units, and command entities will be modeled to the brigade level. OneSAF will
have a variable level of fidelity that supports models and simulations (M&S) domains. It will represent specific
activities of combat, command, control, communications, computers, and intelligence (C4I), combat support, and
combat service support. It will also employ representations of the physical environment and its effect on simulated
activities and behaviors.
Information about the communication environment that must be fed into the Radio
Communication Simulator may take many forms, from simple or complex mathematical models to
simulation or experimental curves. This section discusses three different options to provide these
data to the simulator, namely:

• a single Throughput/Delay (T-D) curve;

• a family of T-D curves; and,
• a Markov process.

A key constraint under which these alternatives must be evaluated is that the results of a
simulation must be reproducible given the same experimental parameters.

7.1 Single T-D Curve

While the notion of network throughput is unclear, both throughput and delay have clear
meanings for a communication link. In network analysis and simulations, T-D curves are often
used as a high level model (or abstraction) of a communication link. The underlying assumption,
that throughput and delay are a full characterization of the communication link, is valid for some
network studies but is not valid in all situations.

Many simulation packages used to produce T-D curves (e.g. OPNET, QUALNET, NS-2)
implement a discrete channel model which assumes independent, randomly distributed errors.
This assumption is often not valid for a wireless channel. In addition to this limitation, a number
of assumptions need to be made to obtain a T-D curve. Some of the required assumptions are:

Network wide:
a. distribution of traffic sources/sinks, i.e. who can originate and receive information;
b. network load and the mix of voice/data in the load (i.e. traffic profiles); and
c. network topology (number of nodes, distance between pair of stations, who is connected
to whom (wired network)).
Each network node:
a. An application traffic model for each workstation
i. distribution of message/packet sizes
ii. distribution (frequency) of message/packet transmissions
iii. distribution of voice/data traffic
iv. voice/data mix
b. (possibly) error-statistics (at least bit error rate) for each individual link between a pair of
stations

An example T-D curve is depicted in Figure 3. For a T-D curve to be valid, a large statistical
sample of the channel conditions (if link error statistics are unspecified) must be exercised. For
the result to be statistically stable, the simulation must be run for a sufficiently long time. The T-
D curve is a statistical curve resulting from a wide range of channel conditions, averaged both
over system parameters and over a long time period. Many of the interesting channel/user
features vanish in this process. Variability, including both good and bad extremes, is lost.
 Figure 3. Illustration of a Throughput/Delay Curve

A second issue is that the throughput/delay curve is valid only for a range/mix of test bed system
parameter values that are consistent with the assumptions made in the derivation of the T-D curve
(message/packet sizes, characteristics of voice/data traffic, number of traffic sources, etc).

When a single T-D curve is used, throughput is the driver. For a given throughput level, all
packets/messages are delayed by a given amount. All other variables have been averaged out (or
eliminated). Since packets are only delayed, never lost, no packet losses can be simulated.

7.2 Family of T-D Curves

A family of throughput/delay curves is the first level of refinement of the T-D curves approach.
With this approach, one can eliminate a strong dependence on a given variable. For example, if
message size were a critical parameter, one could produce a family of curves where each curve is
valid for a (narrow) range of message sizes (see Figure 4).
 Figure 4. Illustration of a Family of Throughput/Delay Curves
A further extension is to have each T-D curve characterized by more than one parameter.
However, the difficulty in such a case is that the size of the family grows quite rapidly with the
number of parameters. Each T-D curve provides a statistical characterization of the channel for a
given set of parameters and a given range of values of these parameters.

As before, each T-D curve should be used within the range of values of system parameters for
which it is valid. As is the case for the single T-D curve approach, if the number of nodes in the
network and/or the network load varies, new T-D curves should be produced.

When a family of T-D curves is used, first a specific (set of) parameter(s) values are used to select
which T-D curve to use, and then throughput is the driver for this set of values. For a given
throughput level, all packets/messages are delayed by a given amount.

7.3 Markov Process

Markov models are used extensively in speech processing, neural networks and other fields.
However, their use for communications applications has developed rather slowly. In the
communication field, they were first brought to light by their use in the derivation of the capacity
of a burst-noise channel [Gilbert, 1960]. In the 1980’s, they were used quite successfully in
studies of HF and satellite channels. Digital error injectors were often based on an underlying
Markov process. More recently, they have been used to model (mobile) wireless channels such as
IEEE 802.11 and GSM [Konrad et al., 2001; Gomez and Campbell, 2000; Golmie and
Mouveaux, 2000; Balakrishnan and Katz, 1998] and even VHF CNR networks [Roman, 1991].
A good summary of models for channels with memory is presented in [Kanal and Sastry, 1978].

The traditional Markov process approach to error modeling is to create a Gilbert-Elliot model
[Gilbert, 1960; Elliot, 1963] (i.e., a two state discrete Markov chain) based upon collected
network traffic traces. Using such a model, one can dynamically generate synthetic network
traces for the network under study and use the traces to simulate the performance of existing and
new network protocols and applications.
 Figure 5. Gilbert-Elliot Channel Model

The Gilbert-Elliott model is characterized by a two-state (N = 2) Markov chain {Cn}n≥0, where Cn

is a random variable associated with the state of the channel at time instant n, with state transition
diagram as shown in Figure 5. Under the Gilbert-Elliott model, the physical channel assumes one
of two states; one state is associated with a low probability of bit error (state 0 or good state),
while the other state is associated with a high probability of bit error (state 1 or bad state), each
having an associated error probability Pe(k), k = 0, 1 and Pe(0) << Pe(1). The state transition
probabilities pij can be derived from the experimental channel data.

Using a Markov Process approach, one can reduce significantly the number of assumptions that
were necessary with the T-D approach. The only piece of information needed to characterize the
Markov model is the error statistics of the link(s). This is used in the derivation of values for the
state transition probabilities. Assumptions regarding message/packet sizes, frequency of
transmission, network load, number of network nodes, etc. are not needed as the process is
applied to each message/packet individually. The Markov process approach results in much finer
granularity of the RCS and a higher degree of fidelity/realism in simulation of the radio channels.

A corollary is that the use of a Markov process introduces an element of randomness in the
simulator. This is due to the fact that one cannot, in theory, predict what is going to happen to a
bit/packet/message until it has gone through the Markov process. However, results are exactly
reproducible due to the fact that, for a given random seed and probability function, a computer
always produces the same outcome.

7.4 Comparison of the Approaches

7.4.1 Stationarity

All three approaches assume that the channel is stationary. Although it is known that the wireless
channel is often not stationary, violating this condition is not as catastrophic as it may seem
because (a) for much of the time the channel will be stationary, and (b) the test bed experiments
are concerned with relative, not absolute results, i.e. one wants to measure the effects of changes
in information management schemes.

7.4.2 Network and Traffic Model Assumptions

The single T-D curve approach makes the most assumptions and averages the results over the
largest set of parameters. It is significantly less precise than the other approaches.

The family of T-D curves approach groups values of one or a set of parameters into ranges that
are used to characterize each T-D curve. It is more precise than the single T-D curve approach.

The Markov process approach does not require any a priori assumptions about the traffic model
except possibly for the voice/data mix. From a network point of view, the distribution of traffic
sources and sinks does not need to be known a priori either. It is therefore the most precise of
the three and introduces a higher level of fidelity (realism).

7.4.3 Network Topology Assumption

Strictly speaking, a Markov process should be associated with each link between nodes since each
of them will in practice have different error statistics. This is a result of the fact that each link
might be of different length (i.e. signal-noise ratio), some stations might be hidden, there might be
different topography between stations, etc. Using a single Markov process is equivalent to saying
that each link between pairs of stations has identical communication characteristics3. These same
considerations apply to T-D approaches.

Whatever approach is selected for the test bed, it is only if the characterization is on a link basis,
i.e., a station pair, that it will be possible to “simulate” these effects and thus to have greater
variability in the consistency of the databases. It is far more feasible to implement characterization
on a link basis with a Markov process than it is with a family of TD curves. This characterization
will be a much more accurate reflection of reality.

7.4.4 Reproducibility of the Results

The T-D curve approaches can be considered deterministic approaches. The Markov process is a
stochastic approach since it involves state transition probabilities between and within states of the
model. However, as noted above, for a given random seed, computers yield the same outcome to
a probabilistic event. All approaches produce results that are strictly reproducible.

8. Conclusions

Quantitative assessment of the impact of information management techniques applied over tactical
radio networks, using the test bed concept discussed in this paper, requires that experimental
conditions and results be highly reproducible and that the sequence of battlefield events be human-
selectable. These characteristics are necessary to allow isolation of the effects of the information
management techniques being tested.

Use of a deterministic battlefield model based on a scripted scenario will provide the required
reproducibility and full control over event sequencing. A stochastic battlefield model, as provided
3
The RCS could be upgraded at a later date to take such effects into account (e.g. take into consideration
geography) if Markov processes are used.
in computer simulation applications like JANUS and ModSAF, produces results that can be made
strictly reproducible if the same random number seed can be employed. However, such a model
will not provide full human control over scenario composition and event sequencing. A
deterministic battlefield model offers clear advantages for the test bed studies.

All three approaches discussed in this paper for the test bed’s communication channel model,
whether deterministic or stochastic, will produce results that are strictly reproducible (provided
that the same random number seed is employed, in the case of the stochastic approach). The
selection of an approach to the communication channel model must be based on another criterion.
This criterion is the model that exhibits the highest degree of fidelity (or realism) to the real
channel behavior while at the same time remaining computationally efficient. The Markov process
approach is clearly superior to the throughput/delay curve approaches in this regard. The
approach offers a much finer granularity and is a much more accurate model of the
communications system. The Markov process approach to the communication channel model
offers a clear advantage for the test bed studies.
9. References

[Gibb and St-Jacques, 1999]A.W. Gibb and J.-C. St.-Jacques. Test bed for the Evaluation of
Battlefield Information Management Techniques Applied to a Low Bandwidth Tactical Wireless
Communications Environment. In Proceedings of the SPIE Thirteenth International Aerosense
Symposium, Digitization of the Battlefield IV Conference, Orlando, Florida, 1999, pp 91-98.

[Banks and Carson, 1984]J. Banks and J.S. Carson II. Discrete-Event System Simulation.
Prentice-Hall International Series in Industrial and Systems Engineering, W.J. Fabrycky and
J.H. Mize, Editors, Prentice-Hall, Inc. Englewood Cliffs, New Jersey, 1984.

[Chamberlain, 1996]S. Chamberlain. Resilient Data Replication Mechanisms for Battle

Command Systems. In Proceedings of the Second International Symposium on Command and
Control Research and Technology, Monterey, CA, 1996, pp 1-11.

[Gilbert, 1960]E.N. Gilbert. Capacity of a Burst-Noise Channel. Bell System Technical Journal,
September 1960, pp. 1253-1265.

[Konrad et al, 2001]A. Konrad, A.D. Joseph, R. Ludwig and B.Y. Zhao. A Markov-Based
Channel Model Algorithm for Wireless Networks. UCLA at Berkeley, Computer Science
Division (EECS), Report No. UCB/CSD-01-1142, May 2001.

[Gomez and Campbell, 2000]J. Gomez and A.T. Campbell. A Channel Predictor for Wireless
Packet Networks. IEEE International Conference on Multimedia and Expo (ICME 2000), New
York, August 2000.

[Golmie and Mouveaux, 2000]N. Golmie and F. Mouveaux. Update on the MAC Coexistence
Modeling Effort. IEEE P802.15 Working Group for Wireless Personal Area Networks,
Document No. IEEE 802.15-00/134r0, May 2000.
[Balakrishnan and Katz, 1998]H. Balakrishnan and R. Katz. Explicit Loss Notification and
Wireless Web Performance. In Proceedings of the IEEE Globecom Internet Mini-Conference,
November 1998.

[Roman, 1991]P.A. Roman. System Effectiveness Model for Combat Radio Networks. MSc
Thesis, Royal Military College of Canada, Dept. of Engineering, Kingston, Ontario, July 1991.

[Kanal and Sastry, 1978]L.N. Kanal and A.R.K. Sastry. Models for Channels with Memory and
Their Applications to Error Control. Proceedings of the IEEE, Vol. 66, No. 7, July 1978,
pp 724-743.

[Elliott, 1963]E.O. Elliot. Estimates of Error Rates for Codes on Burst-Error Channels. Bell
Systems Technical Journal, Vol. 42, September 1963, pp.1977-1997.