We study sensor-actuator networks, extensions of sensor networks that consist of nodes that both monitor and interact with the environment. In particular, we focus on the evaluation of average causal effect within such networks. We describe a distributed algorithm that enables individual actuator nodes to determine the probable consequences of local action on the global environment and hence decide if such action is conducive to achieving the aims of the network. Our approach represents the relationship between actuation and sensor measurements using a causal graph, and applies a distributed expectation-maximization algorithm to estimate the average causal effect of actuation. We evaluate the effectiveness of our approach through simulations that examine the benefits of including side-information regarding possible event outcomes.
|