a learning-based mac for energy efficient wireless sensor ......a learning-based mac for energy...
TRANSCRIPT
A Learning-Based MAC for Energy
Efficient Wireless Sensor Networks
S. Galzarano1,2, Prof. A. Liotta2, Prof. G. Fortino1
(1)University of Calabria, Italy (2)Eindhoven University of Technology, Netherlands
Outline
• WSN & Machine learning
• Learning-based MAC
• Simulation results
• Conclusion & Future work
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 2
Challenges in WSN
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 3
WSN
Wireless Ad Hoc Medium: - unreliable, asymmetric
or unidirectional links - restricted broadband
Topology changes and mobillity: - Mobile sink and/or nodes - failing nodes - new node joining
Harsh environment: - no physical access to
network once deployed - nodes failure
Resource limitations: - battery - processing - memory
Challenges in WSN
Communication Stack Application level
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 4
Clustering
Routing and neighborood
management
Medium Access Control
Physical Layer
Data Processing
Data Collection
Security
Event and target
detection
Medium Access Control
• Protocol layer providing a multiple access control mechanism on a shared communication medium.
• A MAC protocol for WSN should use a radio wake-up/sleep scheduling for: – Energy saving
– Reduce collisions (and then also energy and latency)
– Reduce idle listening periods
– Maximizing throughput
– Minimizing latency
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 5
Research Objective
• Using Machine Learning to improve MAC performance in terms of energy efficiency, throughput and latency.
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 6
Machine Learning in WSN
Machine Learning paradigms have been successfully adopted to address various challenges
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 7
- Can adapt and operate efficiently in dynamic environments.
- Disover important correlation in sensor data
- Support more intelligent decision-making and autonomous control
Reinforcement
Learning
Decision
Tree
Genetic
Algorithms Swarm
Intelligence
R. V. Kulkarni, A. Forster, and G. K. Venayagamoorthy,
“Computational Intelligence in Wireless Sensor Networks: A Survey,”
Communications Surveys Tutorials, IEEE, vol. 13, no. 1, pp. 68 –96, quarter 2011.
.......
Reinforcement Learning
• Usually first choice when solving complex distributed problems in WSNs.
• Trial and error: learning by interacting with the environment: – Learning agents
– Pool of possible actions
– Goodness of actions
– A reward function
– Select one action
– Execute the action
– Observe the reward
– Correct the goodness of the executed action
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 8
• The achieved total reward (Q-value) of taking a specific action at a given state is computed using:
Q-Learning
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 9
• The achieved total reward (Q-value) of taking a specific action at a given state is computed using:
Q-Learning
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 10
new Q-Value old Q-Value
learning constant
Immediate reward old Q-Value
Proposed Q-Learning based MAC
Adapt Q-learning to a radio wake-up/sleep scheduling
• Learning agent Each node in the network
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 11
Proposed Q-Learning based MAC
Adapt Q-learning to a radio wake-up/sleep scheduling
• Learning agent Each node in the network
• State Slot sk in a the frame f
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 12
frame
t
slot sk
Proposed Q-Learning based MAC
Adapt Q-learning to a radio wake-up/sleep scheduling
• Learning agent Each node in the network
• State Slot sk in a the frame f
• Possible actions Radio ON/OFF, for each slot
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 13
frame
t
slot sk
Proposed Q-Learning based MAC
Adapt Q-learning to a radio wake-up/sleep scheduling
• Learning agent Each node in the network
• State Slot sk in a the frame f
• Possible actions Radio ON/OFF, for each slot
• Goodness of actions Q(sk)
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 14
frame
t
slot sk Q(sk)
Reward
Reward signals (per slot)
• Received packets +
• Succesfully transmitted packets +
• Over-heard packets -
• Expected packets -
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 15
Simulation results
• Castalia / OMNET++
• Comparison with other 2 different RL-based MAC
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 16
Simulation results
• Linear, star and mesh topologies
• Packets inter-arrival time between 1 and 10 seconds
• Max throughput is between 20 and 200 byte/sec (200 bytes length payload)
• Performance metrics: throughput, latency, energy efficiency
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 17
Liu, Z., Elhanany, I.: RL-MAC: A reinforcement learning based MAC
protocol for wireless sensor networks. International Journal of
Sensor Networks 1, 117–124 (2006)
Simulation results
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 18
Simulation results
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 19
Simulation results
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 20
Simulation results
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 21
Simulation results
• A nodes-to-sink communication pattern has been used.
• 2 pkt/sec
• Multipath ring routing
• Performance metrics: latency, packet delivery, energy efficiency
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 22
Mihaylov, M., Tuyls, K., Nowé, A.: Decentralized learning in wireless
sensor networks. In: Taylor, M.E., Tuyls, K. (eds.) ALA 2009. LNCS
(LNAI), vol. 5924, pp. 60–73. Springer, Heidelberg (2010)
Simulation results
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 23
Mihaylov, M., Tuyls, K., Nowé, A.: Decentralized learning in wireless
sensor networks. In: Taylor, M.E., Tuyls, K. (eds.) ALA 2009. LNCS
(LNAI), vol. 5924, pp. 60–73. Springer, Heidelberg (2010)
Conclusions
• A Q-Learning approach has been successfully employed for a self-adapting MAC layer on WSNs;
• Simulation results show that it outperforms others RL-based MAC
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 24
Future work
• Ongoing work: – implementation on real sensor platforms;
– extensive experiments with varying real deployment.
• Dynamically update both frame length and slot number on the basis of the network traffic.
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 25
Thank you!!!
Any Questions?
Stefano Galzarano, Antonio Liotta, Gincarlo Fortino 26
A Learning-Based MAC for Energy Efficient Wireless
Sensor Networks
S. Galzarano, Prof. A. Liotta, Prof. G. Fortino
University of Calabria, Italy
&
Eindhoven University of Technology, Netherlands