[Ground-station] Task for FCC TAC AI/ML Safe Uses sub-working-group

[Michelle Thompson]

Greetings all,

This is a task with a very soft deadline of this Wednesday 18 May, but I'm
putting it out to the list for two reasons. First, you all should see what
we're doing at the TAC, and second, it's a really cool assignment and I
know some of you will have valuable opinions and contributions to add.

The AI/ML SWG has given us an assignment to produce a "hot take" on
Bandwidth. Where are we at (with a US-centric view) in terms of using the
spectrum efficiently and well, where do we expect to end up in terms of
Bandwidth occupancy, and what can AI/ML do to contribute to the "Bandwidth
question"? What role does AI/ML play with respect to choices about
Bandwidth?

Our counterpart on this assignment at the FCC is Kambiz Rahnavardy, and
he's super helpful and informed.

Here's the draft so far (included below the bar line).

We're looking for any contributions to any part of this, edits, comment,
critique, but especially some numbers or cases or specific examples. Kambiz
has a CBRS and C/S band focus, so he's working on some content there.

-Michelle Thompson


-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-



Bandwidth Report 18 May 2022
FCC TAC AI/ML Safe Uses SWG
Kambiz
Michelle
et al (your name here)

Bandwidth is the frequency or span of frequencies used for a particular
communications function.

Bandwidth is also widely used to mean the rate of data delivered through a
system. Datarate or throughput is closely related to and limited by the
occupied frequency range of the channel, service, or transmission, but
these two terms are not synonymous. When we talk about the frequency range
occupied by a signal, we will use Bandwidth. When we talk about the amount
of data delivered by a signal per unit time, we will use Throughput.

Bandwidth, along with the signal-to-noise ratio, determines how close we
are to achieving channel capacity.

Channel Capacity = Bandwidth * log2(1 + SNR) bits/sec

The channel capacity limits our ability to detect and correct errors in
digital signals. It’s a limitation that is related to the physics of
entropy. Due to decades of productive work, there are a variety of error
correcting codes that produce spectral efficiencies and therefore
Throughput very close to this channel capacity limit. Two examples of very
high-performance codes are Polar Codes, used terrestrially in 5G, and LDPC
codes, used in space within DVB-S2/X protocols.

A high Bandwidth signal does not necessarily produce a high Throughput data
stream. In the case of a very low signal-to-noise ratio, the Throughput may
be very low compared to the occupied Bandwidth.

There are a variety of constraints and incentives in the design of
communications systems. Bandwidth can be found as both an input parameter
and as an output quantity.

We are almost never completely free to set the Bandwidth in a
communications system. There are regulatory, theoretical, and practical
limits. Regulations may directly limit Bandwidth or may limit the
parameters we use to calculate Bandwidth. Broadband signals generally
require high-performance components to transmit and receive. High
performance usually means high cost. We may want to have more Bandwidth,
but we may simply not be able to afford the components required.

We commonly think of communications design as maximizing Throughput. There
are obvious economic benefits for being able to sell the maximum achievable
capacity to customers. All else being equal, a system that allows 100
simultaneous customers within a geographic area will produce more revenue
than one that can only support 60.

Throughput has balancing requirements. For example, for a mobile device,
power consumption is a hard limit. Higher Throughput and broader Bandwidths
consume more power. Trade-offs between performance and power consumption
depend on multiple factors that may include device cost, acceptable market
price, usage patterns of the target demographic, carrier requirements,
customer expectations and more.

There are IoT devices that operate at extremely low Throughput in very
modest Bandwidths on very low power. There are devices that can remain in
the field reporting telemetry for multiple years on a small battery. In
other words, Bandwidth and Throughput are not blindly maximized. We should
think of Bandwidth (and therefore Throughput) as things we almost always
optimize in the presence of many other constraints and factors. AI/ML has a
powerful role in communications system design because it can determine
optimal solutions in complex search spaces that would simply take too long
to converge using other methods.

Bandwidth is almost always fixed in our communications products. The vision
of dynamic spectrum allocation and cognitive radio, with reconfigurable
radios that perfectly fill in the sea of the noise floor with efficient
digital signaling is often projected as requiring Bandwidth flexibility.
Very few radios are capable of Bandwidth agility. The most relevant
practical reason is that new Bandwidths mean new filters, and filters are
expensive. Changing the Bandwidth on the fly means changing filters on the
fly. Polyphase filter banks are a solution, but agility (maneuverability)
comes at a cost, which is paid either in instability or massive amounts of
parallelized hardware. Polyphase filter banks can be generated on the fly
by AI/ML or they can be pre-rolled. Either way, flexibility with respect to
Bandwidth is costly. Where the AI/ML is done, at this point in time,
doesn’t provide much traction outside of the lab or highly specialized and
very expensive applications. If we want the ability to modify Bandwidth to
adapt to channel or market conditions, then we have to pay in terms of
current consumption, component performance requirements, and complexity.
AI/ML can and does attack and address the component problem today, but
AI/ML is not producing affordable communications systems architectures at
this time. Our theoretical systems architectural reach exceeds our
component level grasp. This is very likely something that will change.
AI/ML assisted component improvements will lead to power consumption
reduction and component level performance. That’s the bottleneck.

In other words, the technical debt incurred with AI/ML cognitive radio
systems can exceed what any company or country is can currently afford. For
example, we can consider the Joint Tactical Radio System (JTRS). The goal
of this program was to replace existing legacy radios in the American
military with a single set of software-defined radios that could be
reconfigured in the field. Even without autonomous AI/ML assisted
functionality, the large size and weight necessary to support the desired
adaptability and flexibility were judged to be much too large to deploy.

AI/ML has a large role in designing better (fixed Bandwidth) communications
systems because the role of AI/ML in areas such as microwave component
design is rapidly growing. The continuing march of miniaturization is a
large factor in how radios like JTRS will eventually be affordable to
field.

What do we currently manage to do in terms of bandwidth occupancy? This
question is nearly impossible to answer compactly, as there’s an enormous
landscape of services, channels, products, and protocols. There are some
similarities across the spectrum (pun intended). As we know, Bandwidth is
almost always fixed in deployed systems and this mandate often comes from
outside the engineering design cycle. A cellular handset company can’t
arbitrarily change the Bandwidth allocated to the cellular service, even if
there were compelling engineering reasons. Changes to deployed device and
system Bandwidth are almost always very expensive. Obtaining more Bandwidth
can be done by innovating and investing in the current spectrum allocation
and/or taking spectrum from others. Both (engineering refactoring and
regulatory efforts) are complex processes where AI/ML is already playing a
role or soon will.

AI/ML applied to the regulatory process itself, in order to produce
optimized spectral Bandwidth allocations and policies, can be an enormously
valuable input. It can also be a highly destructive one. An AI/ML model
trained only on commercial data would annihilate every other use of the
spectrum. These non-commercial uses are valid and are necessary conditions
for the generation and justification of successful companies in the first
place. Educational, non-profit, small company, amateur, scientific, and
unlicensed value is easy to zero out in a spreadsheet that only honors
price to earnings ratio of large publicly traded companies.

AI/ML is only as good as the data it saw in training.

What do we aspire to in terms of bandwidth occupancy? At this point, we
aspire to optimize Bandwidth and Throughput to best achieve communications
goals. Those goals may be conflicting. This is why we have an FCC. The FCC
allocates Bandwidth to produce activity that is judged to have positive
social, economic, political, and security benefits for the United States.

At this point, the only actors that can achieve this balance is humans.
AI/ML does not yet replace the human in the regulatory loop. The training
data required to perform a regulatory role has not been defined, tagged, or
cleaned. AI/ML can and should be part of the process because of the
enormous power it has to cut through properly defined search spaces.

What can we currently manage to do in terms of data rate or throughput?

What do we aspire to in terms of data rate or throughput?

Where can AI/ML “help”? (Have to define who is being helped, really. Some
of these effects are highly negative to consumers while being highly
beneficial to large companies).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openresearch.institute/pipermail/ground-station-openresearch.institute/attachments/20220516/f4f2fce8/attachment.html>