Knowledge-assisted evolutionary task scheduling for hierarchical multiagent systems with transferable surrogates

Task scheduling is a primary step of a hierarchical multiagent system (HMAS) before solving tasks, presenting significant challenges due to its NP-hard complexity and variable-size decision space with different numbers of decision variables. This variability arises because a key decision is determining the number of agents to deploy, which directly affects the dimension of the decision vector. Evolutionary algorithms (EAs) have been widely adopted in addressing the task scheduling problem for HMAS due to their ability to solve NP-hard problems. However, applying conventional fixed-length EAs to such problems often necessitates techniques like expanding the decision space, which negatively impacts search efficiency. Meanwhile, the evaluations of the candidate solutions need physics-based simulations with complex dynamics, which require high computational costs. To solve the HMAS task scheduling problem efficiently, our approach leverages domain knowledge by a genetic programming framework alongside a knowledge-data dual-driven surrogate, which avoids searching in expanded decision spaces and facilitates low-cost evaluation. Notably, the proposed surrogate model can be easily transferred among different task settings, further decreasing the computational load in deploying the HMAS in real-world applications. The effectiveness of the proposed algorithm is validated through extensive simulations on an unmanned ground vehicle/unmanned aerial vehicle (UGV/UAV) cooperation system, showcasing superior efficiency and efficacy. Moreover, the proposed algorithm is also validated in a real-world multi-robot system, further demonstrating the efficacy and efficiency of the method, as well as the transferability of the proposed surrogate model.