Cognitive Radio Systems

A radio that carries a formal model of its own domain, reads policy as data, and can refuse a request that would break its local regulation — then counter-propose. Demonstrated over the air on real radios, and we say exactly how far that demonstration goes.

The Challenge: Radios Cannot Pre-Agree on Everything

Why not just XML, or a JSON schema?

Because a schema fixes the shape of a message and says nothing about its meaning — and an inference engine cannot act on shape. The 2009 paper reaches this verdict on a real competitor — the Functional Description Language (FDL), developed in the E2R (End-to-End Reconfigurability) effort, an XML language expressive enough to state composition, clocks, latency, deadlines and bandwidth — and the verdict is worth quoting in full, because it is the ready-made answer to the objection:

“However FDL (as it stands now) is not a formal system. FDL is defined in terms of XML … Since FDL does not have a formal semantics, it cannot be used by an inference engine for automatic derivation of facts about FDL descriptions.”

The same finding is reached independently, twice more in this body of work, from two different directions. In the link-adaptation work: “XML-represented information cannot be processed by an inference engine since XML lacks formal semantics. Therefore, cognitive radios cannot reason about an XML representation unless explicit information interpretation procedures are written beforehand.” And in the spectrum-consumption-model work, on the XML schema in which those models are actually written: “any modifications to the XSD for SCMML would require rewriting the interpreting code.” That is the whole problem in one line. The meaning of an XML document lives in the program that reads it, so every change to the vocabulary is a change to software — which is exactly the thing you cannot do to a radio that is already deployed and already flying.

And an ontology alone is not enough either

We are not selling OWL as a solvent. The 2009 paper's most valuable results are impossibility results, and the first one lands on OWL itself. OWL can represent a composite component — but it cannot pin down which subcomponent connects to which port, and so it cannot separate the things that are quadrature modulators from the things that merely have the same parts:

“it is impossible to construct an OWL description that would capture all components that are considered to be quadrature modulators and none of the components that are not considered quadrature modulators”

The root cause is stated just as plainly — “This means OWL lacks the capability of expressing the composition of relations” — and the fix is the architecture we have used ever since: rules on top of the ontology. In the paper's summary, “we showed that rules complement OWL and allow us to describe architectures of types of components.” The ontology carries the shared vocabulary and its axioms; the rules compose relations, do the case analysis, and call out to procedural code when the answer needs arithmetic rather than logic.

And the honest cost, which the same paper concedes before anyone else can: “The worst case time complexity of inference in most of the declarative languages discussed in this paper falls in the complexity class termed undecidable.” That is precisely why our production stack does not run unrestricted logic. It runs OWL 2 RL — in the flagship's words, “less than OWL DL but significantly more than propositional logic”, with the complexity “within the range of the decidable.” The expressivity ladder is real, and you must choose a rung you can afford to compute on a radio.

Our Approach: An Ontology, Policies, and a Reasoner Inside the Radio

The stack has been the same for fifteen years, and it has three layers. A shared ontology in OWL, so two radios mean the same thing by the same term. Policies as declarative rules, so behavior can be changed without touching code. And a reasoner in the radio — BaseVISor, ours — so the radio can work out for itself what an unfamiliar message or an unfamiliar rule obliges it to do.

The ontology: shared vocabulary with axioms attached

The Cognitive Radio Ontology (CRO) covers the basic terms of wireless communication at the PHY and MAC layers. It is not a glossary: it is an OWL model with five top-level classes — Object, Process, Quantity, Value and Unit of Measure — whose distinctions are drawn from the DOLCE foundational work, and in the 2011 paper's own summary it runs to 230 classes and 188 properties. It was developed inside the Wireless Innovation Forum's Modeling Language for Mobility Working Group and approved by the Forum in September 2010 as Description of the Cognitive Radio Ontology (WINNF-10-S-0007).

Its successor, CRO2, was rebuilt on NuVio — the small foundational ontology developed by VIStology and Northeastern — and is published at cogradio.org. To be exact about its status: CRO2 is published research, not a Forum-adopted document, and we do not call it a standard. What matters engineering-wise is the discipline it embodies — a foundational layer, a domain ontology on top of it, and application ontologies on top of that, so that a new radio, a new waveform or a new spectrum model extends the model instead of forking it.

The engine: BaseVISor, and rules in BVR

The reasoner is BaseVISor, VIStology's forward-chaining inference engine. In the 2011 link-adaptation radio it is named outright — “In our implementation, we use BaseVISor as the inference engine. The policy rules are expressed in the BaseVISor syntax (BVR)” — and described exactly as we describe it today: “BaseVISor is a forward-chaining rule engine optimized for handling facts in the form of RDF triples. The engine also supports XML Schema Data Types.”

One design decision in that radio has outlived everything else in it. The optimization math — the objective function the radios were trading against — was not written as logic. It was implemented as a procedural attachment: a Java function invoked from inside a rule, with its result bound back to a rule variable. The rules do the situation logic and the arbitration; the attachments do the arithmetic. That division of labor is what makes an ontology-and-rules system usable in the RF domain at all, where half of every real question is geometry or power, and we say so rather than pretending OWL entailment computes distances.

The knowledge in each radio splits three ways: a T-Box (the ontology's classes and properties), an R-Box (the policies and rules, “describing how to react to different situations”), and an A-Box of facts that exist only while the radio is running. Two radios that share the T-Box and the R-Box are mutually intelligible by construction — which is what lets one of them send the other something it has never seen before and expect to be understood.

The refusal: a radio that says no on its own regulation, and then counter-proposes

The sharpest demonstration of what that buys is a message a radio refuses. In the 2011 work, a receiver measures the link and asks the transmitter to change power, coding, modulation and training-sequence length — the receiver has the meter, the transmitter has the knobs, and neither side has both. Control messages ride in an extensible payload rather than in a protocol-defined header, and they are FIPA ACL messages whose content language is OWL/RDF: “we chose OWL/RDF as the content language because it has the machine interpretable syntax and can be directly processed by the inference engine.” One worked message in the paper carries an entire executable rule as its payload — one radio shipping a rule to another radio.

And then the transmitter refuses:

“If the request conflicts with the local regulations (e.g. the transmitter power is out of the permitted range), then radio A will send a ‘refuse’ to radio B. However, the two radios still have the chance to negotiate with each other until an agreement is met.”

Refuse → call-for-proposal → propose → accept. A regulatory constraint held locally by one radio — a constraint its peer may know nothing about — is resolved by counter-offer instead of by failure. That is the behavior an operator actually needs when the same waveform crosses a border, a coalition boundary, or a range restriction that changed last week.

The refusal — two cognitive radios negotiate their own parameters, and one of them says no. A message-sequence diagram between two software-defined radios that negotiate their own transmit and receive parameters. The asymmetry is the reason the conversation has to happen at all: Radio A is the transmitter and holds the knobs — transmit power, training period, and the equalizer parameters m and v — together with its own local regulatory limit on transmit power; Radio B is the receiver and is the only side that can measure the resulting link quality, the measured signal-to-noise ratio mSNR. Neither side alone has both the knobs and the meter. Radio A streams an image to Radio B; B measures mSNR and opens the exchange with a FIPA ACL Query-Ref asking A for its current settings. A agrees, and answers with an Inform-Ref reporting transmit power 0 dB, m equal to 7, v equal to 2 and a training period of 30. B runs its policy, computes what it wants, and sends a Request that A change its transmit power to minus 0.5 decibels. A refuses — not because it disagrees, but because minus 0.5 decibels falls outside the transmit-power range its own local regulation permits. The constraint lives inside Radio A; Radio B was never shipped with it and never had to know it. Rather than failing, B sends a Call-for-Proposal, and A counter-offers: it proposes minus 0.7 decibels instead. B accepts the proposal, A changes its transmit parameters, and A reports Inform-Done. The messages are FIPA ACL performatives whose content language is OWL/RDF, so each radio's reasoner can read them directly; a message can even carry an executable BVR rule that the peer's BaseVISor engine fires on receipt. The honest boundary: this ran on real hardware — GNU Radio with USRP1 radios in the 2.3 to 2.9 gigahertz band, an image streamed while the two radios were physically walked apart — but the paper reports no numbers at all from that run: no timing, no overhead, no convergence, no success rate. The paper's quantitative results, the 66, 55 and 36 percent objective-function reductions, come from a separate MATLAB simulation on an emulated Rayleigh channel, not from the radios. The Refusal Two radios negotiate their own transmit and receive parameters — in OWL, over FIPA ACL — and one of them says no. Neither side alone has both the knobs and the meter. That is why there has to be a conversation. Radio A — the transmitter HOLDS THE KNOBS transmit power · training period · m · v … and its own local regulatory limit Radio B — the receiver HOLDS THE METER equalizer taps M · N1 · N2 the only side that can measure mSNR THE NEGOTIATION refusal is not failure B measures mSNR B runs the policy, computes the optimum A changes its transmit parameters The constraint lives inside Radio A. B never had to know it. B was never shipped with A’s regulatory limits — and it did not need to be. data — an image, streamed continuously Query-Ref PowdB = ? · m = ? · v = ? · trainPeriod = ? Agree Inform-Ref PowdB = 0 · m = 7 · v = 2 · trainPeriod = 30 Request change PowdB to −0.5 dB Refuse −0.5 dB is outside the transmit power my local regulation permits Call-for-Proposal Propose PowdB = −0.7 dB instead — a counter-offer A is allowed to honor Accept-Proposal Inform-Done What is actually inside these messages FIPA ACL envelope 22 performatives · sender · receiver · conversation id Content language: OWL / RDF machine-interpretable — each reasoner reads it directly A message can carry an executable BVR rule one radio ships a rule to its peer; BaseVISor fires it there This ran on real hardware — GNU Radio + USRP1 radios, 2.3–2.9 GHz, an image streamed while the two radios were physically walked apart — but the paper reports no numbers from that run at all: no timing, no overhead, no convergence, no success rate. The paper’s quantitative results — the 66% / 55% / 36% objective-function reductions — come from a MATLAB simulation on an emulated Rayleigh channel, not from the radios. Li, Moskal, Kokar & Brady, “An Implementation of Collaborative Adaptation of Cognitive Radio Parameters Using an Ontology and Policy Based Approach,” Analog Integrated Circuits and Signal Processing 69(2–3):283–296, 2011.
Figure 1. The refusal. Radio B measures the link and asks radio A to drop transmit power to −0.5 dB; radio A holds a local regulatory limit that B knows nothing about, and refuses. The exchange does not fail — B issues a call-for-proposal, A counter-proposes −0.7 dB, B accepts, both radios reconfigure and the link continues. The messages are FIPA ACL performatives; the content inside them is OWL/RDF, read by a BaseVISor reasoner on each end against a shared Cognitive Radio Ontology and a shared rule base. Sequence from Li, Moskal, Kokar & Brady, Analog Integrated Circuits and Signal Processing 69(2–3), 2011 — work done at Northeastern University with VIStology’s BaseVISor as the engine. The 2011 system negotiates; the 2021 policy-interpreter system (covered under Policy-Based Spectrum Sharing) does not, and the two must not be confused.

It ran on real radios — and here is exactly how far that goes

The 2011 loop was closed over the air, not only in simulation: “To further assess our ontology and policy approach, we implemented the link adaptation on GNU/USRP radios.” Two USRP1 software-defined radios with daughterboards operating from 2.3 to 2.9 GHz, a BaseVISor reasoner on each end, an image streamed from one to the other, and the experimenters physically walking the radios apart to perturb the channel. The reported outcome is a working closed loop in both directions — the radios backed off when the link was better than it needed to be, and spent power when it was not.

Now the boundaries, because a technical evaluator will find them anyway:

  • The percentages are simulation. The three policy sets that trade link margin against power efficiency — 66%, 55% and 36% reductions in the objective function, at costs of −1.83 dB, −0.83 dB and +0.09 dB of measured SNR respectively — come from a MATLAB study on an emulated Rayleigh multipath channel, over eight data messages. They are not hardware results, and we never present them as hardware results. What they do show is real: the same radio software, running different declarative policies, produced three different, principled operating points.
  • The hardware run reports no numbers at all. No latency, no control overhead, no convergence time, no success rate. The authors say so themselves and name the measurement as future work: “how much benefit can we get from the use of this approach versus how much overhead it imposes on the communication link in terms of time delay and spectrum efficiency?” Their own word for what they demonstrated is feasibility.
  • It is two radios on a bench, and it is link adaptation — not spectrum sharing. Nobody senses an incumbent; no spectrum is shared. This work supplies the machinery of policy-governed radio behavior. The spectrum-sharing result comes later, and it is the next section.
  • The byline is Northeastern University. The 2009 and 2011 papers were written at Northeastern and funded by DARPA (the 2011 work through System Planning Corporation); VIStology did not hold that work and does not claim it. Jakub Moskal, then a VIStology research scientist, is a co-author of the 2011 paper, and the engine in that radio was VIStology's BaseVISor. That is the accurate claim, and it is enough.

In the Rooms Where the Language Was Written

VIStology's president, Dr. Mieczyslaw M. Kokar, co-chaired the SDR Forum's Modeling Language for Mobility (MLM) Working Group — the group tasked with producing an integrated set of ontologies for the radio sub-domains and a formal language able to express their policies — and was a member of the IEEE 1900.5 Working Group on policy language and policy architectures for dynamic spectrum access. Both roles are stated in his author biography in the Proceedings of the IEEE.

The strongest evidence is not ours, though. It is the Wireless Innovation Forum's own committee outbrief of 14 March 2014, which records, verbatim:

“Mitch Kokar, representing VIStology, led a joint meeting of the Modeling Language for Mobility work Group with IEEE P1900.5 standards group … An overview of the P1900.5.1 draft standard was given by Mitch Kokar. The overview included most of the submission by MLM/Winn Forum of August 2013.”

A standards body, in its own minutes, naming the company and putting its president in front of the draft policy-language standard. That is the proof point, and it is documented at the Forum's own site.

The other artifact of that work is the ontology itself: the Cognitive Radio Ontology was approved as a Wireless Innovation Forum work product — WINNF-10-S-0007 — in September 2010. 230 classes, 188 properties, covering the PHY and MAC vocabulary of wireless communication, and the substrate everything above is built on.

Now the fence. VIStology did not author IEEE 1900.5, and did not author IEEE 1900.5.2. We co-authored peer-reviewed papers about those standards, with the working group's own leadership among the co-authors — which is a real credential and a different one. We have never chaired the IEEE 1900.5 working group. CRO2 is not a Forum-adopted standard. And VIStology was not a performer on DARPA SSPARC. If you have read a spectrum-sharing capability statement that blurs any of those lines, it was not this one.

Proof Points

Over the air

A Radio That Refused — and Counter-Proposed

Two GNU Radio / USRP1 software radios at 2.3–2.9 GHz, a BaseVISor reasoner on each end, a shared 230-class OWL Cognitive Radio Ontology, control messages in FIPA ACL with OWL/RDF as the content language. One radio asked the other to change transmit power; the other refused on its own local regulatory limit, and the pair counter-proposed until they agreed. An image streamed while the radios were physically walked apart, and the loop held in both directions. Boundaries: it is link adaptation, not spectrum sharing; the 66%/55%/36% policy trade-offs are from a MATLAB study, never from the hardware; the hardware run reports no numbers, and the authors name overhead as unmeasured. Northeastern byline, DARPA-funded; the engine was ours.

Selected Publications

Language Issues for Cognitive Radio

Kokar, M.M. & Lechowicz, L.

Proceedings of the IEEE 97(4):689-707 (2009)

View Publication

An Implementation of Collaborative Adaptation of Cognitive Radio Parameters Using an Ontology and Policy Based Approach

Li, S., Moskal, J.J., Kokar, M.M. & Brady, D.

Analog Integrated Circuits and Signal Processing 69(2-3):283-296 (2011)

View PDF

See the full list of VIStology publications →

Related

Dynamic Spectrum Sharing & Cognitive Radio The solution this use case belongs to Policy-Based Spectrum Sharing The sibling use case: the policy engine that returns opportunities Capability Matching The sibling use case: the measured comparison BaseVISor The inference engine inside the 2011 radio Ontologies Including the Cognitive Radio Ontology All Publications The full research record

Put a Reasoner in the Radio

If your radios must act correctly on situations nobody programmed — a peer’s request, a component they lack, a rule they never shipped with — we can help you make the radio read its world as data.

Get in Touch