Q-RC Routing With Compression Using Q-learning
P. Beyens, M. Peeters, K. Steenhaut and A.Nowe presented an interesting approach for learning the best aggregation paths in a WSN.
Routing with Compression in Wireless Sensor Networks: a Q-Learning Approach, presented at the Fifth European Workshop on Adaptive Agents and Multi-Agent Systems (AAMAS ’05), Paris, France, 2005
Describes the algorithm and presents first results using this approach.
The idea is simple and pretty efficient: every node which is producing data and wants to send it to the sink, has a table of all its neighbors. There is an associated Q-value to every of the neighbors, which is being updated through the lifetime of the sensors. The data packet is always being sent to the link (neighbor) with the highest Q-value. This neighbor is then trying to aggregate data from different sources, sends the new data packet further and sends also a feedback to the issuing node (the source, for example). The issuing node uses this data to recalculate the Q-value for this link. Doing so the best link with the highest data aggregation is being selected for next rounds of routing data packets and only minimal amount of data is being sent through the network.
The authors are not talking about how to compute the data aggregation.
The paper is easily readable and understandable. Unfortunately, the authors do not apply the idea to a scenario of multiple sinks. The approach is data-pushing from the sources to the sink and needs also an initialization for the number of hops from every node in the network to the sink. There is also no mobility.