In multi-agent applications it is often the case that not all information is equally valuable to the missions. For example, in the following figure, there is a network of agents that are trying to infer environmental parameters theta, but only two of them (marked as dark nodes) have important information. Since agents are typically resource limited, it is important to ensure that the resources are spent on getting and conveying valuable information. Back to the example, only information from dark nodes is necessary to inform other agents.
This project focuses on developing effecient algorithms that only gather and exploit information that is relevant to mission performanceVoI-based Distributed Sensing
In this case, the mission is to infer a set of parameters theta, but with limited communication resources. Agents broadcast their information only when the VoI in their measurements exceeds a pre-defined threshold. It is proven that communication cost will asymptotically converge to 0 and error is bounded. The VoI threshold is further adaptively adjusted to better balance between the cost and error. The error is proven to converge to 0 in this adaptive algorithm. The following figure compares the performance of VoI-based algorithms (VoIDS, A-VoIDS) with previously proposed algorithms (HPC, Random Broadcast)
In a system with heterogeneous agents, some are able to perform missions are collect scores, while others can gather information and reduce uncertainty in mission score. A VoI-based framework is proposed that evaluate the information gathering by the improvement on mission scores. The hardware experiment shows that the VoI based method performs better in that it couples information gathering directly with mission scores and heterogeneous agents better collaborate on missions.