Instance-driven evolution of constructive heuristic ensemble for the stochastic resource allocation problem with time windows

This paper investigates the stochastic resource allocation problem with time windows (SRA-TW), which is widely encountered in complex systems. In SRA-TW, the assignment of each resource to each task is limited within a time window, and the task completion is described by a time-dependent success probability, aiming to maximize the total expected reward of tasks. To address diverse SRA-TW scenarios, an efficient and general-purpose solving method is urgently needed. We propose an ensemble of multiple constructive heuristics (CHs), which preserves the computational efficiency of individual CHs and exploits their complementarity for superior overall performance. A three-level instance-driven evolution framework (IDEF) is further proposed, where intractable SRA-TW instances guide the adaptive evolution of the ensemble. At the bottom level, a radial-basis-function-network-based CH (RCH) is designed to construct a decision scheme for each instance rapidly, ensuring feasibility through incremental handling of temporal constraints. At the medium level, an evolutionary meta-optimization algorithm (EMOA) is proposed to simultaneously search for an ensemble of RCHs (E-RCH) capable of solving multiple instances. At the top level, intractable instances are iteratively exploited to drive the EMOA to generate new RCHs. By integrating these RCHs and refining them using historical instances, the E-RCH is progressively enhanced in generalization. Experimental results indicate that the E-RCHs built via IDEF can quickly construct decision schemes with higher expected rewards across various test instances, outperforming state-of-the-art algorithms for related problems.