Submodular Optimization via Reinforcement Learning for Active Control of Sensor Networks

Autonomous heterogeneous sensor networks (HSNs) consist of platforms equipped with various sensors. Platforms and sensors have limited range, so they must make decisions on where and when to point these sensors, what information to share among the cooperating platforms, and how to exploit multi-modal dependencies for accurate detection, tracking, and classification (DTC) of time-critical objects. We posit that cooperating platforms within HSN s that share and fuse multimodal data among themselves (using constrained inter-platform communication) will be more effective in accomplishing their required tasks and have more accurate target detection and recognition. We study a scenario where a group of cooperating platforms, each equipped with multi-modal sensors are collecting measurements of an object and the measurements are transmitted over a bandwidth-limited communication channel to a centralized node for processing. The communication channel presents an information transfer bottleneck as the sensors collect measurements at a much higher rate than what is feasible to transmit over the communication channel. In order to minimize the estimation error at the centralized node, only a carefully selected subset of measurements should be transmitted. We propose a unique synthesis of novel submodular optimization based reinforcement learning (RL) for active sensor control/management. Our approach solves the measurement selection problem by employing a model-free RL with sub modular objective functions, where instead of maximizing the cumulative reward, the goal is to maximize the objective value induced by a sub modular function. We make evident the effectiveness of this approach with simulated and real data (the latter including ESCAPE Government-furnished information which we have been extensively using on our previous efforts).