Page 24 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 7 – Terahertz communications
P. 24

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 7




                respective  positions  optimally  in  order    kernel-based  learning,  a  Gaussian  process  regression  is
          to  route  and  maintain  Tbps  links  among  commu-  studied  in  [100]  for  the  channel  estimation  of  a
          nicating  drones.  In  what  follows,  we  review  some  AI/  UM‑MIMO  multi‑user  system  over  the  THz  band  (0.06‑
          ML‑based  approaches  possibly  implementable  for   10 THz).  It is to be noted here that the AI/ML techniques
          THz‑enabled  drone  networks  for  channel  estimation,   mentioned  above  are  proposed  for  6G  wireless  commu‑
          UM‑MIMO, and Mobile edge computing.                  nications  and  networks  at  sea  level  in  general.  There‑
                                                               fore,  there  will  be  a  need  to  tailor  these  AI/ML
          5.1  Channel estimation                              techniques  speci ically  for  the  THz‑enabled  drone
                                                               networks  and DSNs keeping in view the intrinsic nature
          As  discussed  earlier  in  Section  1,   irst,  the  THz  band  is   of  the  communicating  drones  i.e.,  mobility,  energy‑
          highly affected from absorption loss due to water vapor   constrained  resources etc.
          molecules in the atmosphere, contributing signi icantly to
          the total loss.  Second, the spread loss is also massive at   5.2  UM‑MIMO
          THz frequencies.  Third, the THz band channels are non‑
          stationary, particularly for mobile use cases i.e., hovering   In  the  realm  of  UM‑MIMO,  ML  can  be  employed  in
          drones in our case, where both the Tx and Rx drones will   various  use  cases.  One  instance  is  when  an  existing
          be  mobile.  Hence,  conventional  assumptions  of  quasi‑   model  is  erroneous,  and/or  it  is  only  a  sparse
          stationary  or  stationary  channel  models  may  not  be  ap‑   approximate  of the  actual  model.  Such  an  instance  can
          plied  to  the  THz  band  channels.  More  speci ically,   arise  within  the  linear  channel  models,  where  the
          channel  estimation  over  the  THz  band  becomes  more   non‑linearities  induced  by  certain  practical  circums-
          challenging in the drone scenarios under mobility, where   tances  and  hardware  are  neglected.  Also,  ML  can  be
          precise  Channel  State  Information  (CSI)  is  needed,  e.g.,   utilized  for  improving  the  solu‑  tions  obtained  using
          in  beam‑forming.   Therefore,  the  traditional  channel   approximations  of  the  linear  mod‑ els.  Another instance
          estimation  methodologies  are  required  to  be  revisited   of ML‑based solution in UM‑MIMO is  possible  when  the
          [52].  Overall,  for  reducing  the  complexity  of  the  THz   optimized  solutions  are  computa‑ tionally expensive i.e.,
          channel   estimation,   several   techniques   can   be   not feasible for the state‑of‑the‑art hardware.  Here, ML
          employed  includ‑ ing:  compressed sensing, fast channel   can  be  effectively  utilized  for   inding  suboptimal
          tracking‑based algorithms,  etc.  ML‑based algorithms,  in   solutions  having  less  complexity,  with  some  obvious/
          this  context,  can  be  employed  for  evaluating  the  THz   acceptable lower performance.  Examples in this  context
          band  communication  data  by  anticipating  the  THz   include   channel   estimation,   maximum   likelihood
          signal  loss  in  a  certain  unknown channel.  Consequently,   detection,  etc.  Moreover,  optimum  spectrum  utilization
          various AI or ML‑based algorithms are applicable  to the   in UM‑MIMO can be made possible using machine learn‑
          physical layer of the forth‑coming  6G  wireless  networks   ing techniques [22].
          for  addressing  the  above‑mentioned THz channel model
          and  estimation  [92,  93].  Supervised  Learning  (SL)  [94]   5.3  Mobile edge computing
          can  aid  in  predicting  THz  shadowing  and  path  loss.
          Moreover,  SL  can  be  employed  for localization, channel   Mobile  Edge  Computing  (MEC)  has  recently  emerged  as
          estimation,  interference  management,  etc.  Employable   a technique for 5G networks, in which cloud computing‑
          SL  models  and  algorithms  are  K‑Nearest  Neighbor   like functionalities are processed at the edges of the cellu‑
          (KNN),  Support  Vector  Machine  (SVM),  feed‑forward   lar networks [101].  MEC can equip mobile devices, such
          neural  networks,  and  radial  basis  function  neural   as  drones  with  constrained  processing  capabilities,  to
          networks,  etc.  Various  challenges  related  to  the  THz   hand over their processing tasks to the nearest network’s
          channel  modeling  and  estimation  including  multi‑path   edge. Conversely, within a THz‑enabled DSN, mobile user
          tracking,   interference   mitigation,   node   clustering,   equipment can of load its computationally expensive jobs
          optimized modulation, etc., can be tackled by using Unsu‑   to  the  serving  drones  with  MEC  functionalities.  Low  la‑
          pervised  Learning  (UL)  techniques  [95]  such  as:  Fuzzy   tency systems for sporadic access such as cyber‑physical
          C‑means,  K‑means,  clustering  algorithms,  etc.  Deep   communication  systems  (a.k.a.,  Tactile  Internet)  [102]
          Learning  (DL)  (both  SL  and  UL)  can  be  employed  in   require latencies within sub‑ms for controlling hovering
          many  aspects  of  channel  modeling,  such  as  for  signal   objects (drones in our case).  Such alike systems will also
          detection,  and  estimating  Channel  State  Information   be  a  requirement  for  evolved  industry  4.0  applications
          (CSI).  Techniques  including  Deep  Neural  Networks   [103]. It is predicted that the transport methods over the
          (DNNs),  Recurrent  Neural  Networks  (RNNs),  Convo-  physical layer will be linked with edge computing such as
          lutional  Neural  Net‑  works  (CNNs)  can  be  anticipated   MEC, or  real‑time  cloud  computing  within  the  vicinity
          as  appropriate  candidate  DL  algorithms  [96].  Reinfor-  of  the  communication  network.  The  main  objective
          cement  Leaning  (RL)[97]  can  be  used  for  channel   here  is  to  deliver  resources/solutions  to  the  evolving
          selection  and  tracking,  iden‑  ti ication  of  radio  bands,   IoT  protocols  comprising  of  a  massive  number  of
          selection  of  modulation  modes,  etc.   Appropriate  RL   inter‑connected  devices  with  constrained  energy  and
          models  and  techniques  include  Q‑learning,  fuzzy  RL,   storage  requirements  such  as  in  drone  networks,  as
          etc.   [98].   Finally,  learning‑based  schemes  for  THz   well  with  some  latency requirements.  For  overcoming
          channel  estimation  is  highly  ef icient particularly  over   these  constraints,  global 6G research is moving towards
          higher  dimensions  [99].   As  an  exam    deep     distributed computation techniques (MEC here) [104, 105].
          12                                 © International Telecommunication Union, 2021
   19   20   21   22   23   24   25   26   27   28   29