PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepMIMO: A Generic Deep Learning Dataset for Millimeter Wave and Massive MIMO Applications

Ahmed Alkhateeb

arXiv:1902.06435v1cs.ITeess.SP

TL;DR

mmWave and massive-MIMO machine-learning research needs a common dataset for evaluation, reproduction, benchmarking, and comparison. DeepMIMO addresses this need with Wireless InSite ray-traced channels and adjustable system and channel parameters. The paper describes an example with 18 base stations and more than one million users and demonstrates beam prediction.

  • Problem

    A common, sufficiently large dataset is needed to evaluate, reproduce, benchmark, and compare mmWave/massive-MIMO machine-learning algorithms.

  • Method

    DeepMIMO constructs channels from accurate Wireless InSite ray-tracing data and lets researchers tailor datasets through adjustable system and channel parameters.

  • Results

    The paper presents an example dataset based on 18 base stations and more than one million users and demonstrates its use for mmWave beam prediction.

  • Takeaways & Limitations

    DeepMIMO supports application-specific dataset construction, result reproduction, and algorithm comparison using a common dataset definition.

Abstract

from arXiv · show

Machine learning tools are finding interesting applications in millimeter wave (mmWave) and massive MIMO systems. This is mainly thanks to their powerful capabilities in learning unknown models and tackling hard optimization problems. To advance the machine learning research in mmWave/massive MIMO, however, there is a need for a common dataset. This dataset can be used to evaluate the developed algorithms, reproduce the results, set benchmarks, and compare the different solutions. In this work, we introduce the DeepMIMO dataset, which is a generic dataset for mmWave/massive MIMO channels. The DeepMIMO dataset generation framework has two important features. First, the DeepMIMO channels are constructed based on accurate ray-tracing data obtained from Remcom Wireless InSite. The DeepMIMO channels, therefore, capture the dependence on the environment geometry/materials and transmitter/receiver locations, which is essential for several machine learning applications. Second, the DeepMIMO dataset is generic/parameterized as the researcher can adjust a set of system and channel parameters to tailor the generated DeepMIMO dataset for the target machine learning application. The DeepMIMO dataset can then be completely defined by the (i) the adopted ray-tracing scenario and (ii) the set of parameters, which enables the accurate definition and reproduction of the dataset. In this paper, an example DeepMIMO dataset is described based on an outdoor ray-tracing scenario of 18 base stations and more than one million users. The paper also shows how this dataset can be used in an example deep learning application of mmWave beam prediction.

I. INTRODUCTION

DeepMIMO addresses the lack of a common mmWave/massive-MIMO dataset by combining environment-aware ray-traced channels with adjustable dataset parameters. This supports evaluation, reproduction, benchmarking, and comparison across machine-learning studies.

  • A sufficiently large common dataset is needed to evaluate algorithms, reproduce results, and benchmark and compare solutions.
  • Existing mmWave/massive-MIMO datasets lacked the stated requirements for environment representation and parameterized use.
  • DeepMIMO introduces a channels dataset designed for machine-learning research in mmWave and massive MIMO applications.
  • Ray-tracing data from Wireless InSite capture channel dependence on environment geometry, materials, and transmitter/receiver locations.
  • Researchers can tailor datasets by adjusting parameters controlling system and channel aspects such as antennas, OFDM subcarriers, and channel paths.
  • The dataset is completely defined by its ray-tracing scenario and parameter set, enabling straightforward generation and reproduction.

II. DEEPMIMO DATASET: THE GENERAL FRAMEWORK

The DeepMIMO framework combines a ray-tracing scenario with adjustable parameters and generation code to produce application-specific channel datasets. This design preserves geometry-based channel characteristics while supporting reproducibility and algorithm comparison.

  • Ray-tracing simulations provide geometry- and material-dependent channel characteristics, including correlations across user locations.
  • Researchers tune parameters such as antenna count, array configuration, and subcarrier count to fit their application.
  • The ray-tracing scenario supplies channel parameters for transmitter-receiver pairs distributed across an indoor or outdoor environment.
  • The parameter set S controls dataset generation, providing control over system setup and antenna configuration.
  • The generation code uses scenario channel parameters and S to construct channel matrices and include features such as user locations.
  • Because the dataset is defined by scenario R and parameter set S, researchers can reproduce datasets and compare algorithms on common data.

III. DEEPMIMO DATASET: A DETAILED DESCRIPTION

Dataset generation starts from a selected ray-tracing scenario and its channel parameters, then applies researcher-selected system parameters to construct the DeepMIMO dataset.

  • Researchers choose an available ray-tracing scenario whose transmitter-receiver channel parameters provide the first input to the generation code.
  • The researcher adjusts parameter set S for the target application, after which the code constructs the dataset from the scenario and system parameters.

A. Ray-Tracing Scenarios

DeepMIMO uses accurate Wireless InSite ray-tracing scenarios to capture how channel parameters depend on environment geometry, materials, and transmitter–receiver locations. The example ’O1’ scenario models two streets, an intersection, 18 base stations, and more than one million users.

  • Ray-tracing outputs capture environment geometry, materials, transmitter-receiver locations, path gains, and arrival/departure angles.Wireless InSite provides the channel parameters used as inputs to the dataset generation framework.
  • The ’O1’ scenario is an outdoor layout with two streets and one intersection.The main street is 600m long and 40m wide, while the second street is 440m long and 40m wide.
  • 18 base stations are distributed along both streets in the ’O1’ scenario.Twelve base stations serve the main street and six serve the second street; all are 6m high.
  • 1,184,923 users are placed across three uniform x-y grids in the scenario.The grids cover the streets with 20 cm spacing between adjacent users in the first grid.
  • The scenario models solid rectangular buildings and 60 GHz propagation materials.Buildings use dimensions of 30m × 60m or 60m × 60m, with ITU dry earth and layered drywall materials.
  • Wireless InSite uses a highly accurate 3D propagation model and retains the strongest 25 paths after considering the first four reflections.The simulation shoots hundreds of rays for each transmitter-receiver pair.

B. DeepMIMO Dataset Parameters

DeepMIMO parameters let researchers tailor the generated dataset by selecting active transmitters and users, antenna configurations, bandwidth, OFDM sampling, and channel-path counts. These controls also reduce dataset size when only subsets of the system or channel representation are needed.

  • Active BS and user parameters select which base stations and user-row ranges contribute channels.Selecting subsets focuses the dataset on the application’s required transmitter-receiver pairs and users.
  • Antenna parameters specify uniform-array dimensions and element spacing relative to wavelength.For example, a 16 × 16 UPA in the y-z plane uses num_ant_x=1, num_ant_y=16, and num_ant_z=16; half-wavelength spacing uses ant_spacing=.5.
  • The bandwidth parameter defines system bandwidth in GHz.A 500 MHz system is represented with bandwidth=.5.
  • OFDM parameters specify the number and subset of subcarriers used for channel calculation.Sampling and limiting subcarriers can reduce dataset size.
  • The num_paths parameter selects how many of up to 25 received-power-ordered paths are included.Setting num_paths=3 retains the three strongest paths.

C. DeepMIMO Dataset Construction Code

The DeepMIMO generation code combines ray-tracing channel parameters with the selected dataset parameters to construct channel vectors for active base-station–user pairs and subcarriers. It also exports user locations and makes the dataset reproducible from the scenario and parameter set.

  • The code constructs an M × 1 channel vector for every active BS-user pair on each selected subcarrier.M is the total number of BS antennas, expressed as M=M_xM_yM_z.
  • The channel construction uses array-response vectors for the BS antenna array.The array response is factored into x-, y-, and z-direction response vectors.
  • The dataset includes user locations alongside the OFDM channel vectors.The location p_u=[p_x,p_y,p_z] can serve as an input feature for machine-learning applications.
  • The constructed dataset is determined solely by the ray-tracing scenario R and parameter set S.This two-input definition supports dataset description, reproduction, and algorithm comparison.

D. Structure of the DeepMIMO Dataset

For each active base station and user, DeepMIMO stores channel matrices over selected subcarriers together with user positions. The generated data are packaged in a single MAT file with a cell-array structure for access.

  • Each active BS-user entry contains an M × |K| channel matrix over the specified subcarriers.Each column represents the channel at one selected subcarrier.
  • Each entry also stores the active user’s three-dimensional position vector.The position is accessed as the user location field.
  • The generated outputs are stored in one MAT file named DeepMIMO dataset.mat.The file contains a single cell array called DeepMIMO_dataset.
  • Channel data are accessed through DeepMIMO_dataset{b}.user{u}.channel using active-BS and active-user indices.These indices refer to positions within the selected active subsets, not necessarily the original identifiers.
  • User positions are accessed through DeepMIMO_dataset{b}.user{u}.loc.The location field returns the position vector for the selected active user.

IV. HOW TO USE THE DEEPMIMO DATASET?

DeepMIMO generation requires defining a ray-tracing scenario and parameter set, then following a simple workflow to obtain the tailored dataset.

  • The dataset is generated from a researcher-defined ray-tracing scenario and parameter set tailored to the desired application.
  • The workflow involves downloading the generation code, downloading the adopted scenario’s ray-tracing outputs, and placing the scenario folder in the specified directory.

V. AN EXAMPLE APPLICATION: BEAM PREDICTION

DeepMIMO supports a supervised beam-prediction application that maps omni-received OFDM sequences from multiple base stations to beamforming outputs and achievable rates.

  • The coordinated beamforming model uses uplink omni-antenna pilots received at multiple base stations to predict the best beamforming vector at each coordinating base station.
  • The dataset is completely defined by its ray-tracing scenario and parameter set, supporting dataset reproduction and algorithm comparison.
  • Each supervised-learning data point contains the omni-received OFDM sequence at four base stations and achievable-rate outputs for candidate beamforming vectors.
  • DeepMIMO channels provide the per-base-station, per-subcarrier channel vectors used to construct the model inputs and achievable-rate outputs for each user.
  • The generated data train the coordinated beamforming model, whose performance is evaluated through achievable-rate results versus training-set size.

VII. CONCLUSION

The paper presents DeepMIMO as a ray-tracing-based, parameterized channel dataset for machine learning in mmWave and massive MIMO systems.

  • DeepMIMO constructs MIMO channels from accurate Remcom Wireless InSite ray-tracing data, capturing dependence on scatterer geometry and transmitter/receiver locations.
  • Researchers can generate datasets using adjustable system and channel parameters, including an example scenario with 18 base stations and more than one million users.
  • The paper demonstrates constructing inputs and outputs for a deep-learning coordinated beamforming application using DeepMIMO.
Loading 1902.06435v1…