Machine Learning for Wireless Sensor Networks
Context
Wireless sensor networks (WSN) form an emerging class of networks able to monitor environments with high spatiotemporal accuracy. The network is composed of tiny devices known as wireless sensors or motes, endowed with a microprocessor, a memory, a radio, a battery, and one or more sensors such as temperature, humidity, light or sound sensors. The figure below gives an illustration of a typical wireless sensor platform used in research (TMote), and of an integrated silicon design (Deputy Dust).
The transmission of data from a WSN to an observer raises numerous issues: wireless sensors are constrained by limited resources, in terms of energy, network data throughput, and computational power. The communication module is a particularly constrained resource since the amount of data that can be routed out of the network is inherently limited by the network capacity. Also, wireless communication is an energy consuming task, identified in many situations as the primary factor of lifetime reduction.
Project proposal
The goal of this thesis is to continue the research in the field of reinforcement learning and routing algorithms that has already been performed at our lab. Our main focus is to use learning techniques in order to improve network routing while allowing for application dependent Quality of Service (QoS).
The quality of service is application dependent, and several criteria may be considered for optimizing the network performances: energy consumption, application lifetime, accuracy of the collected data, transmission reliability, latency, …. An important issue in setting the parameters of a network is that QoS criteria are often interdependent: increasing the quality along an axis often leads to decrease the quality along another axis. Improving the network lifetime while limiting the impact on the network latency has been the focus of some of our earlier research (Devillé, 2011).
Machine learning and reinforcement learning provide sets of techniques which allow the successful self-organization of a network. The major advantage of machine learning techniques, over their competitors, is to determine in an almost autonomous manner the best parameters of a system. Examples include self-synchronization and desynchronization of a network (Mihaylov, 2010), the compression and prediction of sensor data (Le Borgne, 2010) and reinforcement learning for energy-aware routing in wireless sensor networks (Devillé, 2011).
The student will be expected to work on a topic related to learning and sensor networks, and to collaborate with researchers of the COMO team. The exact topic will be determined on the basis of the student's motivation.
The student will have the opportunity to experiment and test its solutions thanks to a WSN which has been installed in the VUB greenhouses (on top of the E building). The purpose of the sensor network is to accurately monitor, over space and time, the variations of temperature, humidity, and light. It is composed of 40 prototypical TMote Sky (see picture above). Once installed, the network is expected to periodically collect data from all sensors (every 5 minutes), and to make them available in real-time through Internet (the network will already be running when the student starts).
Contact
Promoter
Supervisor
References
- [1997, book | bibTeX]
- T. M. Mitchell, Machine Learning., , 1997.
- [2009, phdthesis | bibTeX]
- Y.-A. Le Borgne, "Learning in Wireless Sensor Networks for Energy-Efficient Environmental Monitoring," PhD Thesis , 2009.
- [2011, conference | bibTeX]
- M. Devillé, Y.-A. Le Borgne, and A. Nowé, "Reinforcement Learning for Energy Efficient Routing in Wireless Sensor Networks," in Proc. Proceedings of the 23rd Benelux Conference on Artificial Intelligence, Ghent, Belgium, 2011, pp. 89-96.
- [2010, inproceedings | bibTeX]
- M. Mihaylov, Y. Le Borgne, A. Nowé, and K. Tuyls, "Decentralized Reinforcement Learning for Wake-up Scheduling," in Proc. Proceedings of the 7th European Conference on Wireless Sensor Networks, 2010.
