Placement-strategy selection guide

Placement strategies decide where each signal in a scene lands in time and frequency. Choosing the right pair per scene is the difference between a dataset that trains a foundation model to recognize realistic RF environments and one that teaches the model to memorize uniform-random layouts that no real receiver ever sees. This page is written for ML engineers who are PyTorch- and statistics-fluent but not yet RF experts; it maps each of the framework’s 11 shipped strategies to its evidence tier, cites the primary source, and gives a copy-paste default recipe per common scene type. The full 11-strategy catalog with per-strategy pseudocode lives in Scene Composition Algorithm.

Evidence tier per strategy

Every shipped strategy is classified by the strength of the empirical grounding for its behavior. The tiers are defined in How to interpret the evidence tiers below.

Frequency-placement strategies

Strategy	Evidence tier	Primary citation	Modeling target
IIDUniformFreq	TEXTBOOK_STANDARD	Cover & Thomas, Elements of Information Theory Ch 12 (maximum-entropy baseline)	Maximum-entropy carrier draw with no domain prior.
StratifiedFreq	ENGINEERING_PRIOR	(no external citation; standard variance-reduction technique)	Guaranteed spectral coverage across bins for benchmarking.
RealisticDensityFreq	REGULATORY_FACT (centers) / ENGINEERING_PRIOR (weights)	IEEE 802.11-2020; Bluetooth Core 5.4; LoRa Alliance RP002-1.0.5; ICAO Annex 10 Vol IV	Standards-defined channel grids with prior weights.
ClusteredFreq	ENGINEERING_PRIOR	(canonical anchor-plus-jitter model)	Anchor-and-jitter around a handful of protocol-defined centers.
ForcedOverlap	ENGINEERING_PRIOR	(no external citation; adversarial training convention)	Deliberate cochannel collisions for detector stress tests.

Time-placement strategies

Strategy	Evidence tier	Primary citation	Modeling target
IIDUniformTime	TEXTBOOK_STANDARD	Cover & Thomas, Elements of Information Theory Ch 12 (maximum-entropy baseline)	Uniform onset with no cadence prior.
EventRadarPRI	TEXTBOOK_STANDARD	Richards, Fundamentals of Radar Signal Processing Ch 1.4; Skolnik, Introduction to Radar Systems § 3.6	Pulse-repetition-interval cadence for radar.
EventPeriodicBeacon	REGULATORY_FACT	IEEE 802.11-2020 § 11.1.3 (Wi-Fi TBTT); Bluetooth Core 5.4 Vol 6 § 4.4.2 (BLE adv); 3GPP TS 38.213 § 4.1 (NR SSB); ICAO Annex 10 Vol IV § 3.1.2.8 (ADS-B)	Standards-mandated periodic control-frame cadence.
EventBurst	PEER_REVIEWED_RESEARCH	Willinger, Taqqu, Sherman, Wilson 1997, IEEE/ACM TON 5(1), 71–86 doi:10.1109/90.554723	Heavy-tailed Pareto on/off bursts for data traffic.
EventBurstSelfExciting	PEER_REVIEWED_RESEARCH	Hawkes 1971, Biometrika 58(1), 83–90 doi:10.1093/biomet/58.1.83	Self-exciting point process; models packet cascades and retransmit storms where prior events increase the probability of subsequent events (temporal correlation, not just heavy-tailed marginals).
EventFhssHop	TEXTBOOK_STANDARD	Simon, Omura, Scholtz, Levitt 2001, Spread Spectrum Communications Handbook Ch 4	Frequency-hopping dwell schedules.

Default recipes per scene type

Scene	Time strategy	Frequency strategy	Rationale + citation
Mixed ISM 2.4 GHz (Wi-Fi + BLE + data traffic)	EventPeriodicBeacon for control frames + EventBurst (or EventBurstSelfExciting for correlated bursts) for data	RealisticDensityFreq (wifi-2.4ghz + ble manifests)	IEEE 802.11-2020 § 11.1 + Bluetooth Core 5.4 Vol 6 § 1.4.1 for cadence and channels; Willinger et al. 1997 (Pareto) or Hawkes 1971 (self-exciting) for burst distribution
Radar-only	EventRadarPRI	ClusteredFreq (few anchors) or a single fixed center	Richards 2014 Ch 1.4; Skolnik 2001 § 3.6
Cellular (LTE / 5G NR DL)	EventPeriodicBeacon (SSB at 20 ms)	ClusteredFreq around ARFCN centers	3GPP TS 38.213 § 4.1 (SSB periodicity); 3GPP TS 38.104 (NR band ARFCN raster)
Generic (no domain prior)	IIDUniformTime	IIDUniformFreq with min_spacing_hz on the order of the widest expected signal bandwidth	Cover & Thomas Elements of Information Theory Ch 12 (max-entropy baseline)

Copy-paste config recipes

# Mixed ISM 2.4 GHz
scene:
  time_placement: event_periodic_beacon
  frequency_placement: realistic_density
  time_placement_params: {period_s: 0.1024}         # Wi-Fi TBTT default (IEEE 802.11-2020 §11.1.3)
  frequency_placement_params: {taxonomy: wifi-2.4ghz, min_spacing_hz: 5_000_000.0}

# Mixed ISM 2.4 GHz, correlated data traffic (self-exciting Hawkes)
scene:
  time_placement: event_burst_self_exciting
  frequency_placement: realistic_density
  # baseline_rate_hz: mean event rate absent excitation
  # excitation_strength must be < 1 for stationarity (Hawkes 1971)
  time_placement_params: {baseline_rate_hz: 5.0, excitation_strength: 0.6, decay_rate_hz: 20.0}
  frequency_placement_params: {taxonomy: wifi-2.4ghz, min_spacing_hz: 5_000_000.0}

# Radar-only
scene:
  time_placement: event_radar_pri
  frequency_placement: clustered
  # 1 ms PRI is a typical surveillance value; tracking radars run ~100 us,
  # LPI/agile radars ~10 us. See Skolnik 2001 § 3.6.
  time_placement_params: {pri_seconds: 0.001, jitter_s: 0.0}
  frequency_placement_params: {anchors_hz: [9.5e9]}

# Cellular (LTE / 5G NR downlink)
scene:
  time_placement: event_periodic_beacon
  frequency_placement: clustered
  # 20 ms is the default SSB periodicity in 3GPP TS 38.213 § 4.1.
  time_placement_params: {period_s: 0.020}
  # Anchors are ARFCN centers from 3GPP TS 38.104 (example: n78 mid-band).
  frequency_placement_params: {anchors_hz: [3.5e9, 3.6e9, 3.7e9]}

# Generic (no domain prior)
scene:
  time_placement: iid_uniform
  frequency_placement: iid_uniform
  # min_spacing_hz set to the widest expected emitter bandwidth prevents
  # trivially-overlapping draws without imposing a domain-specific channel plan.
  frequency_placement_params: {min_spacing_hz: 20_000_000.0}

How to interpret the evidence tiers

• REGULATORY_FACT cites a standards document; the values are what real devices are required to use. Trust boundary: the standard itself.

• TEXTBOOK_STANDARD is a canonical model in the field, cited to a widely-used textbook. Trust boundary: the model is the community-agreed default; parameter values may still need tuning for a specific scenario.

• PEER_REVIEWED_RESEARCH cites a specific peer-reviewed paper. Trust boundary: the paper’s findings. If your scenario differs from what the paper measured (band, protocol, deployment density), results may differ.

• ENGINEERING_PRIOR is a sensible convention widely used but with no direct empirical grounding for the specific parameter values. Trust boundary: use as a starting point, tune if you have measurement data.

Weights in the shipped channel-grid manifests

The RealisticDensityFreq strategy loads a per-taxonomy manifest that ships with the package. Manifests carry channel centers (from standards) plus channel weights (mostly engineering priors). This table makes that split explicit.

Manifest	Centers grounded in	Weights grounded in
wifi-2.4ghz.json	IEEE 802.11-2020 Cl 19	Engineering prior (canonical 1/6/11 triplet from DSSS-bandwidth arithmetic; not measurement-backed)
wifi-5ghz.json	IEEE 802.11-2020 Cl 17 (UNII bands)	Engineering prior
ble.json	Bluetooth Core 5.4 Vol 6 § 1.4.1	Engineering prior (channels 37/38/39 upweighted because they are the advertising channels, Bluetooth Core 5.4 Vol 6 § 4.4.2)
lora-us915.json	LoRa Alliance RP002-1.0.5 Table 2-3	Engineering prior
lora-eu868.json	LoRa Alliance RP002-1.0.5 Table 2-7	Partial regulatory grounding (RP002-1.0.5 § 2.4.3 duty-cycle rules imply mandatory channels are used ~2x optional)
adsb-1090mhz.json	ICAO Annex 10 Vol IV § 3.1.2.8	Trivial (single channel)

Weights are prior distributions, not empirical measurements. Kokkoniemi & Lehtomäki (2012) CROWNCOM published measured 2.4 GHz occupancy that could replace the Wi-Fi weights; the framework’s BaseChannelPlanSource plug-in surface (see the RealisticDensityFreq strategy subsection) lets you ship measurement-backed channel plans as an alternative source without touching the default.

Related work in RF-ML datasets

For readers coming from the ML side who want to compare rfgen’s placement choices to prior work:

• RadioML 2016.10A / 2018.01A (O’Shea, Corgan, Clancy 2016 arXiv:1602.04105; O’Shea, Roy, Clancy 2018 doi:10.1109/JSTSP.2018.2797022) ship single-signal-per-example datasets with uniform-random offsets. rfgen’s IIDUniformFreq + IIDUniformTime combination reproduces this.

• TorchSig / Sig53 Wideband (Boyd et al. 2022 arXiv:2207.09918) produces multi-signal wideband scenes with uniform frequency draws (rejection on spectral overlap threshold) and uniform onset. This is the closest prior-art peer.

• Sionna (Hoydis et al. 2022 arXiv:2203.11854) provides link-level and ray-traced simulation with transmitters at user-provided coordinates; there is no built-in placement-policy library. Cite Sionna for propagation, not placement.

• Stochastic-geometry cellular (Andrews, Baccelli, Ganti 2011 IEEE TCOM 59(11), 3122–3134 doi:10.1109/TCOMM.2011.100411.100541) is the canonical Poisson-point-process model for cellular base-station placement. rfgen does not implement this today; it is future work.

Future work

Documented placement models rfgen does not yet implement:

• Semi-Markov FHSS (hop-dwell with state-dependent duration; Simon et al. 2001 Spread Spectrum Communications Handbook Ch 4).

• Poisson-cluster / Neyman-Scott frequency placement for large device-swarm scenes (Neyman & Scott 1958 doi:10.1111/j.2517-6161.1958.tb00272.x).

• Cellular Poisson-point-process placement (Andrews, Baccelli, Ganti 2011, above).

See also

• Scene Composition Algorithm - full 11-strategy catalog with per-strategy pseudocode.

• Concepts / Scenes - conceptual overview.

• Reference / Config Schema § Selecting placement strategies - YAML surface.