Przemek Wozniak, Los Alamos National Laboratory; Svetlana Riabova, Los Alamos National Laboratory; Hassan Hijazi, Los Alamos National Laboratory
Keywords: Space Domain Awareness, multi-agent systems, sensor networks, optimization
Abstract:
The convergence of autonomous robotic hardware and distributed sensor ecosystems is revolutionizing surveillance technology and observational science. Next generation systems for Space Domain Awareness will deploy networks of intelligent agents acting in the environment, requiring a mix of collaborative and individual behavior to accomplish their mission. While numerous approaches are being explored, this long-term vision presents often unique challenges. Autonomous Tip & Cue systems are often the only way to “catch” and characterize emerging events of interest before it is too late to respond or there are too many events. The key missing piece is efficient coordination that can quantify the utility of information and optimize the overall outcome. Despite rapid progress on multiple fronts, this problem remains one of the main barriers to a wider adoption of such systems.
Here we present a solution based on the idea of using a market-place model to represent interactions within distributed multi-agent sensor networks designed to coordinate sensor ecosystems with diverse sensing capabilities, but limited resources, and no strict centralized control. We then apply this approach to the problem of Space Traffic Management with globally distributed sensors tracking a large population of space objects and minimizing the overall uncertainty and collision risk, while simultaneously optimizing individual objectives of all agents. We also demonstrate the gain in total utility from coordinated data collections compared to uncoordinated behavior. Typically, agents have limited information toward satisfying their objectives, which creates a strong incentive to trade with other agents. However, at any given time there is no obligation to participate in the market, allowing agents to retain autonomy and pursue their own goals.
The system consists of agents that own private sensors and a coordination authority (CA), that operates the exchange market for trading observations, but does not own any sensors. Sensor trade creates possibilities for the agents to observe some objects of interest and engage additional capabilities, which are not accessible from their own sensors. The CA, however, does not have any research agenda and is acting as a market regulator by setting the exchange prices to reduce the total uncertainty associated with the population of space objects. The presence of CA allows agents to avoid sharing private information (observation utilities) with each other. While the market approach is introduced to enable effective collaboration between autonomous agents, it can also be used in systems with centralized control as a way of managing complexity.
The proposed market system features a two-level hierarchy, which can be viewed as a bi-level game where the CA acts as a leader by setting the price for exchanged observations and the agents (followers) react to that pricing decision by committing to certain exchanges based purely on their interests. In this kind of game, each player solves their own optimization problem, the leader tries to generate the best outcome while accounting for the followers’ response, and the followers act in the space parameterized by the leader maximizing their individual benefit. The knowledge of agents’ utilities allows the CA to reconstruct their individual optimization problems and solve a bi-level problem predicting the response to the pricing policy.
A key aspect of bi-level optimization is that given a leader’s decision, the followers always react optimally, making these models a powerful tool to capture the relationship between agents in a market. Certain caveats are also discussed. The model features a mechanism to ensure fair observation pricing and agents are guaranteed that the resulting sensor assignment benefits their research programs. The proposed approach requires solving a mixed-integer optimization problem, which provides a globally optimal solution and is known to be NP-hard. While the optimal solution can in principle be found, in practice it is often preferred to obtain a near-optimal solution in a reasonable time.
Date of Conference: September 19-22, 2023
Track: Space Domain Awareness