Optimizing Sensor Network Fusion for Improved Localization Accuracy in Dec-POMDPs

This paper considers joint optimization of sensing and communication in a distributed sensor network that uses a shared wireless channel to relay sensor measurements to a fusion center. In the scenario we consider, the sensors’ measurements are used to estimate the distance to a vehicle that is moving through the environment according to a Markov process. The sensor network is tasked with providing the fusion center with sufficient information to estimate the vehicle’s location accurately, but because the channel is shared, collisions occur when multiple sensors transmit during the same interval. At the same time, highly accurate measurements from a single sensor may not be sufficient to localize the vehicle because a single sensor can only determine range and not the full position in two- or three-dimensions. Thus, techniques are needed for each sensor to determine whether it should transmit to the fusion center in each observation interval, with the cooperative goal of minimizing the location error estimate at the fusion center. The resulting system can be modeled as a decentralized partially observable Markov decision process (Dec-POMDP). We use multi-agent reinforcement learning (MARL) to approximately solve the Dec-POMDPs for the sensors by quantizing the beliefs and applying distributed tabular Q-learning. We compare the performance to two benchmark algorithms, and the results show that our MARL algorithm is able to achieve significantly better localization performance than the benchmark algorithms in most cases.