RLinWiFi –
optimizing network
thorughput with
Deep Reinforcement
Learning
Witold Wydmański, AGH
The bottleneck
Sources:
freepik.com
pixabay.com
Presentation plan
 Reinforcement learning in IoT
 General benefits in wireless communication
 A small trick in our sleeves
IoT. Wired networks – channel access
problem
Sources:
https://www.itrelease.com/2019/06/what-is-bus-topology-with-example/
http://spoat.jebrp.proe.hendil.mohammedshrine.org/basic-lan-wiring-diagram.html
Router
IoT. Wireless networks – channel
separation
2.4 GHz
IoT. Wireless networks – channel
separation
2.4GHz 2.5GHz 2.4GHz 2.5GHz
Markov Decision Process
MDP elements Tic-tac-toe Channel selection
State
Action
Reward
MDP elements Tic-tac-toe Channel selection
State Board
Action Drawing a sign
Reward Win
MDP elements Tic-tac-toe Channel selection
State Board Previous attempts
Action Drawing a sign Choosing a channel
Reward Win Lack of collision
Sources:
https://www.zabkat.com/blog/tic-tac-toe-learning.htm
http://ashleywang.me/machine-learning/2017/06/05/RL-Tictactoe.html
Deep Q-network
Sources:
https://medium.com/@carsten.friedrich/part-3-tabular-q-learning-a-tic-tac-toe-player-that-gets-better-and-better-fa4da4b0892a
2.4GHz
2.5GHz
IoT. Wireless networks – channel
separation results
IoT. Energy Harvesting
Sources:
https://anilkrpandey.wordpress.com/2020/02/09/energy-harvesting-and-wireless-power-transfer-antenna-for-iot-applications/
IoT. Energy conservation
Sources:
Energy-Efficient IoT Sensor Calibration with Deep Reinforcement Learning, A. ASHIQUZZAMAN et al.
Reinforcement learning based multi-access control and battery prediction with energy harvesting in IoT systems
WiFi. Medium Access Control
Sources:
https://blogs.arubanetworks.com/industries/real-world-examples-and-discussion-around-wifi-channels/
WiFi. Distributed Coordination Function
DIFS
DIFS 1 Transmission
Delayed access
Draft and
backoff
8
DIFS 15
Frame
CW max
CW min
CW
Draft range
0 15 1023
WiFi. Reinforcement Learning for CW
optimization
Sources:
Contention Window Optimization in IEEE 802.11ax Networks with Deep Reinforcement Learning – W. Wydmański, S. Szott
https://www.freecodecamp.org/news/an-intro-to-advantage-actor-critic-methods-lets-play-sonic-the-hedgehog-86d6240171d/
MDP elements CCOD
State Observed collision probabililty
Action Choosing a CW value
Reward Throughput
CW max
CW min
CW
0 15 1023
RLinWiFi. Centralized Contention
window Optimization with DRL
RLinWiFi. Results
Basic scenario Convergence scenario
The trick
Sources:
https://www.nextplatform.com/2017/03/21/can-fpgas-beat-gpus-accelerating-next-generation-deep-learning/
Conclusions

RLinWiFi: Optimizing network throughput with Deep Reinforcement Learning