You are on page 1of 15

Deep Learning Aided-BLER Estimation for

Maximum Likelihood Detection


in MIMO Systems

Le Van Thuan
Advisor: Prof. Kyungchun Lee
Seoul National University of Science and Technology
Communications and Signal Processing Lab.
Group meeting (16th July, 2020)
Contents
 Introduction
 Review of general concepts
– Physical Layer Abstraction
– Effective SINR Modeling
– Effective SINR Mapping Methods
– Effective SINR Modeling for MLD
 Deep Learning-based BLER Estimation
 Simulation Results

2
Introduction
 Link adaptation denotes the matching of the appropriate modulation and coding
scheme (MCS) to the current conditions on the radio link.
 System level simulator evaluates the channel condition, i.e., effective SINR or BLER
and sends to higher layer.

3
Physical Layer Abstraction
 The goal of PHY abstraction is to accurately predict link level performance in system simulation.

Link level simulation

Full receiver with demod


Channel Model
and decoder BLER1

Transmitted packets,
MCS, and SNR
Effective Eff. SINR to BLER
Channel Model Equalizer SINR mapping, using a
computation reference curve
BLER2

PHY abstraction

 The PHY abstraction must be designed such that BLER1=BLER2.


– Can be obtained through applying reasonable “Effective SINR computation” block and a matching BLER reference curve.

4
Effective SINR Mapping (ESM)
 In
  general, effective SINR can be calculated as
follows:

1 N

 
SINRn
SINReff    
1

N

n 1    

– is the post-processing SINR at the subcarrier


– is number of data subcarrier used in OFDM system.
– is Effective SINR Mapping function.
– and are tunable parameters.
 Effective SINR is used to map the equivalent BLER
from AWGN performance curve for a particular
modulation and coding scheme (MCS).

5
Effective SINR Mapping Methods

6
Effective SINR Mapping for MLD
  ESM requires SINR for each subcarrier , then it is only valid for linear
receiver, i.e., ZF or MMSE receivers.
 For non-linear receiver, i.e., MLD or SD, it is hard to compute SINR for
each subcarrier due to complex mathematical relation.
 Difficulty in parameters optimization problems.
 Therefore, the accuracy of this method is low.

7
Deep Learning based Effective SINR Estimation
 Motivation: a Deep Learning methods is proposed to analyze the relation between
the input features (channel matrix and SNR) and BLER (or effective SINR).

Link level simulation

Full receiver with demod


Channel Model
and decoder BLER1

Transmitted packets,
MCS, and SNR

Deep Neural
Channel Model BLER2
Network with DL

PHY abstraction

8
Deep Learning based Effective SINR Estimation
 Regression Problem

H R
Deep Neural Deep Neural
Network with DL BLER Network with DL BLER
SNR SNR

9
Deep Learning based Effective SINR Estimation
  Classification Problem: BLER is quantized by bits
– Use Multi-Class Binary Classification with sklearn Deep Learning Library

  𝑏1   𝑏1
H R
Deep Neural
Network with DL
  ⋮ Deep Neural
Network with DL
  ⋮
SNR   𝑏𝑄 SNR   𝑏𝑄

10
Simulation Results
 Mean Absolute Error  System Parameters

Parameters Values
Channel Model Rayleigh fading
Modulation QPSK
 Mean Bias Error Channel Coding Turbo
Coding Rate 1/2
Packet size 1024 bits
Samples 5284 ( MIMO)
16947 ( MIMO)
 10% Accuracy Criterion

11
Simulation Results
  MIMO

MAE Min MAE MBE Min MBE 10% Accuracy Max 10% Accuracy
DNN with 4 bits 0.2082 0.0921 0.0281 0.0921 60.05% 67.69%
DNN with 5 bits 0.1709 0.0456 -0.0208 0.0456 73.32% 83.76%
DNN with 6 bits 0.1529 0.0224 -0.0450 0.0224 80.26% 91.98%
DNN with 7 bits 0.1442 0.0110 -0.0568 0.0110 83.14% 95.61%
DNN with 8 bits 0.1404 0.0058 -0.0621 0.0058 84.22% 97.01%

12
Simulation Results
  MIMO

MAE Max MAE MBE Max MBE 10% Accuracy Max 10% Accuracy
DNN with 4 bits 0.1992 0.0924 0.0366 0.0924 60.26% 68.18%
DNN with 5 bits 0.1627 0.0456 -0.0120 0.0456 73.71% 84.24%
DNN with 6 bits 0.1453 0.0228 -0.0355 0.0007 79.91% 91.91%
DNN with 7 bits 0.1365 0.0113 -0.0474 0.0113 83.05% 95.75%
DNN with 8 bits 0.1324 0.0059 -0.0530 0.0059 84.49% 97.46%

13
Simulation Results

  MIMO   MIMO
In log scale In log scale
0 0

-0.5 -0.5

-1 -1
Estimated BLER

Estimated BLER
-1.5 -1.5

-2 -2

-2.5 -2.5

-3 -3
-3 -2.5 -2 -1.5 -1 -0.5 0 -3 -2.5 -2 -1.5 -1 -0.5 0
True BLER True BLER

14
15

You might also like