Effective Control of Leader-Follower Networks
Abstract
We will study networks of of autonomous agents which come in two types which we will refer to as leaders and followers, respectively. The main difference is in the level of intelligence available to each agent: leader agents are able to do orders of magnitude more computations per unit time as compared to followers. This division is motivated by recent advances in large-scale machine learning, which have resulted in neural networks capable of human-level performance in image recognition, language translation and understanding, recognition of human actions, as well as strategic game playing. We envision future networks of multi agent systems, composed of ground vehicles and UAVs working in concert, which are able to perform these tasks in real time on the battlefield. Such networks could make strategic decisions in fulfillment of a global objective, automatically adapt based on image, video, and speech collection, and notify human supervisors via voiced transmissions. Unfortunately, recent advances in machine learning enabling these capabilities have relied on an extremely large amount of computational power, orders of magnitude more than a UAV or an agile robot/vehicle can handle. This motivates a network architecture which divides agents into two types: "bulkier leaders who can run state-of-the-art learning algorithms and the more agile followers who are guided by the leaders. In our model, followers respond to leaders via the nearest neighbor influence mechanism: each follower is influenced by its neighbors, which are in turn influenced by their neighbors, and so forth. By contrast, the leaders are not influenced by the followers. At a high level, this model allows the leaders to make decisions and choose trajectories with the expectation that the nearest neighbor interactions will cause the rest of the network to "fall into place in terms of fulfilling the mission objective. Our main concern with this architecture is how a small group of intelligent leaders can steer a large multi agent network; in particular, we wills study how many leaders are needed and where they should be positioned in the network for such control to be possible. A prototypical example is the V formation, which allows one leader to effectively steer a network of followers in a desired direction. Focusing on the problems of coverage control and formation steering, we will design interaction graphs and control strategies allowing a small group of leaders to steer the network in fulfillment of changing coverage and formation objectives. Our technical approach is based on "control of networks framework. We will show that effective control of leader-follower networks can be achieved by resolving open problems regarding actuator placement in complex networks. Unfortunately, previous work by the PI has shown that such problems are, in full generality, computationally intractable, so we will focus on approaches which hold under technical assumptions that can be expected to hold in real-life scenarios...
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1810072
Entities
People
- Alexander Olshevsky
Organizations
- Army Contracting Command
- Boston University
- United States Army