Channel Gain Prediction for Cooperative Multi-Agent Systems

Licentiate thesis, 2016

In a cooperative multi-agent system (MAS), agents communicate with each other using the wireless medium. As agents move in the environment in order to fulfill the MAS' higher level task, their location changes and so does the wireless communication channel they experience. To enable a successful coordination, it is paramount for the agents to retain connectivity among themselves. In order to achieve this, the availability of explicit channel knowledge for the MAS' future configuration is needed. Since the agents determine their location from sensors, any expected residual location uncertainty for the MAS' future configuration will have an implication on the channel knowledge. For this reason, a computationally attractive yet accurate method to predict the wireless ad-hoc communication channel for any configuration and location uncertainty of the agents is needed. In this thesis, we employ Gaussian processes (GPs) for learning of channel model parameters and for channel prediction at arbitrary (unvisited) transmitter (TX) and receiver (RX) locations. In an indoor measurement campaign, we investigate the ad-hoc wireless communication channel and its properties with respect to path-loss and shadowing from obstacles. We derive a suitable GP model, where we incorporate spatial correlation of communication links caused by shadowing. The effectiveness of our approach in a cooperative MAS is demonstrated, where the bit error rate (BER) among the agents' communication links is minimized. Furthermore, we extend our GP framework allowing to make distributed predictions using a consensus scheme. We found that the incorporation of location uncertainty into channel prediction allows to outperform approaches where this is neglected. The incorporation of location uncertainty at both, the TX and the RX location, leads not only to robust estimates of the underlying channel parameters, but also to realistic channel predictions with respect to the agents' true location uncertainty. Applied to a cooperative MAS, we see that the BER and BER uncertainty can be significantly reduced. Finally, with a distributed channel prediction, we observe a trade-off between computation complexity and accuracy of prediction. Natural extensions of our GP channel prediction framework could include distributed parameter learning and efficient methods to handle a high number of measurements.