Adaptive reconstruction mechanism for enterprise collaborative networks based on multi-agent reinforcement learning
Year: 2026
Type: article
Abstract: Enterprise collaborative networks must reorganise as markets shift, partners fail and demand changes. This study develops an adaptive reconstruction mechanism in which each firm acts as an autonomous agent that adds, terminates or reweights collaborative ties within a bounded candidate set. The Multi-Agent Deep Q-Network with Prioritized Experience Replay (MADQN-PER) combines attention-based coordination, a dynamic learning rate and rewards that mix firm-level and network-level objectives. Graph operations have explicit enterprise meanings, authorising parties and cost components; action masks enforce feasibility before selection. Across synthetic networks of 50 to 200 nodes, the reported efficiency gains range from 34.7% to 41.2%, with 78.3% performance retention under 15% random node failures. Attention and the shared reward are associated with the largest component differences in composite score. A calibrated supply-chain simulation reports a 41.7% reduction in delivery time and a 53.2% robustness gain. Quantitative comparisons are restricted to the five baselines with documented configurations. Conservative bounds derived from ten-run summaries distinguish gains in efficiency, robustness and adaptability from the reconstruction-cost trade-off. The evaluation schedule supports finite-budget comparisons but does not establish convergence times.
Source: Scientific Reports
Authors: Jinxia Peng, Menglin Li, Yiming Li
Institutions: Guizhou University of Commerce, Woosong University
Language: English
Subfield: Management of Technology and Innovation
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7225227921