TY - JOUR
T1 - Instance-adaptive coevolutionary learning of constructive heuristics for multi-objective stochastic resource allocation with preferences
AU - Zhang, Jingyu
AU - Xin, Bin
AU - Liu, Bo
AU - Xu, Meng
AU - Zhang, Jia
AU - Wang, Qing
N1 - Publisher Copyright:
© 2026 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
PY - 2027/2/1
Y1 - 2027/2/1
N2 - In time-critical settings such as rapid-response operations, decision makers must allocate limited resources to tasks with uncertain completion outcomes under multiple conflicting objectives. With decision-maker preferences, this becomes a multi-objective stochastic resource allocation problem with preferences (MOSRAP), where the goal is to quickly obtain high-quality solutions in the region of interest (ROI) rather than approximate the entire Pareto front. Constructive heuristics can rapidly build feasible solutions, but their fixed construction criteria often encode a single implicit preference, limiting their adaptability across preferences and problem instances. To address this, this paper proposes IACoL, a coevolutionary training approach that learns preference-conditioned portfolios of constructive heuristics offline for fast online decision making. IACoL represents construction criteria using a parameterized fuzzy evaluation system and trains heuristic portfolios with adaptively generated MOSRAP instances through coevolution to improve cross-instance generalization. Once trained, the portfolios directly construct ROI-oriented solutions for new instances. Experiments on diverse MOSRAP instances show that IACoL outperforms state-of-the-art preference-based algorithms in solution quality, ROI coverage, computational efficiency, and generalization.
AB - In time-critical settings such as rapid-response operations, decision makers must allocate limited resources to tasks with uncertain completion outcomes under multiple conflicting objectives. With decision-maker preferences, this becomes a multi-objective stochastic resource allocation problem with preferences (MOSRAP), where the goal is to quickly obtain high-quality solutions in the region of interest (ROI) rather than approximate the entire Pareto front. Constructive heuristics can rapidly build feasible solutions, but their fixed construction criteria often encode a single implicit preference, limiting their adaptability across preferences and problem instances. To address this, this paper proposes IACoL, a coevolutionary training approach that learns preference-conditioned portfolios of constructive heuristics offline for fast online decision making. IACoL represents construction criteria using a parameterized fuzzy evaluation system and trains heuristic portfolios with adaptively generated MOSRAP instances through coevolution to improve cross-instance generalization. Once trained, the portfolios directly construct ROI-oriented solutions for new instances. Experiments on diverse MOSRAP instances show that IACoL outperforms state-of-the-art preference-based algorithms in solution quality, ROI coverage, computational efficiency, and generalization.
KW - Coevolutionary learning
KW - Constructive heuristics (CH)
KW - Multi-objective optimization
KW - Preference
KW - Stochastic resource allocation (SRA)
UR - https://www.scopus.com/pages/publications/105054170802
U2 - 10.1016/j.eswa.2026.134636
DO - 10.1016/j.eswa.2026.134636
M3 - Article
AN - SCOPUS:105054170802
SN - 0957-4174
VL - 334
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 134636
ER -