Model based multi-agent numeric planning system for spacecraft

Planning for spacecraft is a complex problem, with diverse behaviours of heterogeneous spacecraft systems and varied mission requests. Traditional space mission planners treat a spacecraft as a single entity. In fact, by taking each subsystem as an agent, one spacecraft is a loosely coupled multi-agent system in which subsystems complete missions by cooperation. For such a system, the multi-agent planning method is more efficient. Also, there are durative actions and numeric variables in the problem, which raise numeric constraints like temporal and resource constraints. These factors make spacecraft planning a numeric planning problem, which is not well studied in multi-agent planning problems. The proposed system aims to combine multi-agent planners' efficiency and the numerical variable processing capacity of classic spacecraft planners. In this paper, we develop a model-based multi-agent numeric planning system for spacecraft. A multi-agent numeric domain modelling language (MANDML) is used to express spacecraft knowledge and the mission requests with temporal and resource constraints. This knowledge will be put in model files as the planner's input. This model-based problem description makes the planner domain independent. We define privacy and public information for each agent to preserve agents' privacy and increase communication efficiency. Instead of relying on a management agent to coordinate agents, we design a decentralized multi-agent planning architecture that allows agents to plan in parallel. A dynamic agent interaction graph (DAIG) based on public information is proposed to coordinate agents under the decentralized architecture. We propose a novel multi-agent plan-space planning method. The planning process of each agent starts from an initial private partial plan. Agents search for a complete plan by flaw resolution. When there are public flaws that an agent cannot resolute by itself, it will ask for help through DAIG. And numeric constraints handling algorithms based on multi-agent constraints network (MACN) are inserted into the planning process to propagate numeric constraints through agents. Finally, a comparison with a classic spacecraft planner EUROPA highlights the advantage of this system.