Research on Multi-Agent Collaborative Decision-Making Algorithm for Supply Chain Management

Addressing the key challenges of fuzzy credit allocation, low exploration efficiency, and insufficient robustness in multi-node collaborative decision-making in supply chain management, this paper proposes a hybrid local-global credit allocation multi-agent collaborative decision-making algorithm (HGA-MADDPG). This algorithm introduces a hierarchical graph attention mechanism to dynamically represent the state of the supply chain network topology. It quantifies the contribution of individual actions to sub-chain objectives and system-level indicators through local and global credit networks, respectively, and designs an adaptive fusion weight based on marginal returns to dynamically balance local and global credit. Furthermore, an adversarial disturbance and resilient training architecture is constructed, including modeling three types of disturbances: demand mutation, node failure, and transportation delay, as well as adversarial agent injection, a dynamic environment replay buffer, and a two-stage training strategy. In a baseline scenario of a four-level supply chain and a dynamic environment driven by real data based on SCDL and WSN, compared with eight baseline algorithms, experimental results show that HGA-MADDPG achieves a total cost reduction rate of 26.2%, a service level improvement rate of 42.8%, and a stockout rate controlled at 3.2%. In the extreme scenario of triple perturbation, the cost deviation rate (29.6%) and recovery time (58 hours) are significantly better than existing methods. It still maintains a cost reduction rate of 21.5% in a 120-node ultra-large-scale supply chain. Ablation experiments and scalability analysis further verify the effectiveness of each core module.

Summary

Main Finding

The paper proposes HGA-MADDPG, a multi-agent collaborative decision-making algorithm for supply chain management that combines a hierarchical graph-attention state encoder, dual (local + global) credit networks with an adaptive fusion weight, and a coordinated exploration + adversarial/resilient training architecture. In experiments (four-level benchmark driven by real data), HGA-MADDPG outperformed eight baselines: it reduced total cost by 26.2%, improved service level by 42.8%, kept stockout rate at 3.2%, and showed much stronger robustness under extreme triple perturbations (cost deviation 29.6% and recovery time 58 hours superior to existing methods). It still achieved a 21.5% cost reduction in a 120-node ultra-large scenario. Ablation and scalability tests attribute gains to each core module.

Key Points

Problem addressed: ambiguous credit assignment, inefficient collaborative exploration, and insufficient robustness for multi-node, non-stationary supply chains under disturbances.
State representation: hierarchical graph attention network (two layers)
- Local encoding layer: one-hop upstream/downstream attention per node.
- Global fusion layer: a virtual global node aggregates local encodings to provide broadcasted global context.
- Dynamic attention update: attention scores adjusted by a supply–demand matching error term so the GAT responds to disturbances (e.g., shortages, node failures).
Dual credit architecture:
- Local credit network (per agent) estimates contribution to local sub-chain objectives (inventory cost, local shortages) and provides immediate feedback.
- Global credit network evaluates joint actions’ effect on system-level metrics (total cost, delays, service penalties).
- Adaptive fusion weight: a small network outputs each agent’s weight between local vs global credit, updated based on marginal global contribution and local advantage (so agents dynamically emphasize local or global guidance as appropriate).
Collaborative exploration:
- Learnable collaborative exploration matrix M couples exploration noise across agents (initialized by topology; adapted by gradient on global exploration reward), producing positively/negatively correlated exploration among neighbors to speed discovery of coordinated policies.
Adversarial & resilient training architecture:
- Disturbance models: demand mutation, node failure, transportation delay.
- Adversarial agent injection to create extreme scenarios during training.
- Dynamic environment replay buffer that prioritizes challenging samples.
- Two-stage adversarial–cooperative training strategy to balance average performance and robustness.
Experimental outcomes:
- Compared to eight baselines (including standard MADDPG-based methods), HGA-MADDPG gives large improvements in cost, service, stockouts.
- Robustness: significantly better recovery and smaller cost deviation under compounded perturbations.
- Scalability: retains substantial benefits at 120 nodes.
- Ablation studies demonstrate contribution of hierarchical GAT, dual-credit design, adaptive fusion, collaborative exploration, and adversarial training.

Data & Methods

Algorithmic basis: extension of MADDPG (centralized training, decentralized execution) with:
- Hierarchical GAT encoder producing per-agent representations (local + broadcasted global).
- Local Q_i^loc networks trained with local rewards (inventory and shortage penalties).
- Global Q_glb network trained on system-level reward (normalized operational cost, delays, service penalties).
- Fusion network φ_i mapping agent state to a sigmoid weight λ_i; combined critic Q_i_total = λ_i·Q_i_loc + (1−λ_i)·Q_glb.
- Collaborative exploration: noise ε = M η + ζ, with M learned by gradient-based objective on global exploration reward.
- Training with soft target updates and experience replay; adversarial data generation and prioritized replay for disturbances.
Evaluation setup:
- Primary benchmark: a four-level supply chain (supplier → factory → warehouse → retailer) with dynamics driven by real data (paper states "based on SCDL and WSN").
- Baselines: eight comparative algorithms (including vanilla MADDPG variants and other multi-agent / heuristic methods).
- Metrics: nine metrics across cooperative efficiency, decision quality, and algorithm performance (examples: total cost, service level, stockout rate, cost deviation under disturbance, recovery time).
- Stress tests: extreme (triple) perturbation scenarios and ultra-large 120-node supply chain simulations.
- Ablation: remove/disable modules (hierarchical GAT, local/global credits, fusion, collaborative exploration, adversarial training) to measure marginal impact.
Key reported quantitative results:
- Baseline scenario: total cost −26.2% vs baselines; service level +42.8%; stockout rate 3.2%.
- Extreme triple-perturbation: cost deviation 29.6% and recovery time 58 hours better than compared methods (paper reports “significantly better”).
- 120-node scenario: cost reduction maintained at 21.5%.
- Ablation/scalability results indicate each module meaningfully contributes to performance and robustness.

Implications for AI Economics

Mechanism design for decentralized networks: The hybrid credit architecture operationalizes a practical separation between local incentives and system-level objectives. This is directly relevant to economic mechanism design in networks where local agents have private information and local goals but system efficiency requires coordination.
Internalizing externalities: The global credit network and adaptive fusion weight act like a learned internalization of cross-node externalities (e.g., upstream capacity congestion caused by downstream over-ordering). Such learned “taxes/subsidies” via credit signals could inform automated market rules or pricing schemes to reduce inefficiencies (bullwhip effects).
Robustness value in supply-chain resilience: Quantified reductions in cost deviation and recovery time under shocks show the economic value of robust, learned coordination policies — pertinent for firms and policymakers investing in resilient supply chain AI.
Scalability and adoption: Demonstrated gains at 120 nodes indicate potential applicability to large real-world supply networks, suggesting meaningful aggregate welfare and cost-savings if widely adopted. However, engineering, privacy, and institutional integration costs remain nontrivial.
Labor and market structure: Automation of decentralized coordination may shift skill demands (more emphasis on systems/AI oversight) and could change bargaining power along supply chains if coordination reduces information asymmetries.
Cautions and next steps for economists:
- External validity: results are based on simulated benchmarks driven by specific datasets (SCDL, WSN); field trials and out-of-sample real-world deployment are needed to validate economic outcomes.
- Strategic behavior & incentives: paper assumes cooperative agents following learned policies. In real markets, independently owned firms may have misaligned objectives; integrating strategic incentives (contracts, transfer pricing) into the learning framework is important.
- Data and privacy: centralized training accesses full-state information during training; practical deployment must address data sharing, privacy, and regulatory constraints.
- Policy leverage: regulators could consider how such AI coordination tools interact with competition policy — improved coordination may provide efficiency gains but could also facilitate collusion if used across competing firms.

Overall, HGA-MADDPG offers a promising technical approach for improving decentralized coordination and robustness in complex supply chains; from an AI-economics perspective, it provides a concrete mechanism for aligning local agent behavior with system-level objectives and highlights avenues for further work on incentives, deployment, and empirical validation.

Assessment

Paper Typedescriptive Evidence Strengthlow — Results are produced in simulated supply-chain environments driven by datasets (SCDL, WSN) and extensive baselines/ablation tests, but there is no randomized field trial, natural experiment, or quasi-experimental design to establish causal effects in real-world operations; external validity is therefore limited. Methods Rigormedium — The paper proposes a sophisticated algorithm (hierarchical graph attention, local/global credit networks, adaptive fusion, adversarial/resilient training), evaluates against eight baselines, performs ablation and scalability tests, and stresses the model under multiple disturbance scenarios; however, evaluation is simulation-based with limited discussion of hyperparameter sensitivity, statistical uncertainty, real-world deployment constraints, or observational heterogeneity. SampleExperimental evaluation on simulated multi-node supply-chain environments: a baseline four-level supply chain driven by dynamic real-data traces (reported as SCDL and WSN), extreme scenarios with three simultaneous perturbations (demand mutation, node failure, transport delay plus adversarial agents), and scalability tests up to a 120-node supply chain; comparisons to eight baseline algorithms and ablation experiments reported. Themesproductivity innovation GeneralizabilitySimulation-based results may not translate to operational real-world supply chains with unmodeled constraints, Proprietary or underspecified datasets (SCDL, WSN) limit replication and external validation, Performance may depend on specific network topologies and parameter settings used in experiments, Modeled disturbances may not capture the full variety or timing of real-world shocks, No evidence on integration with human decision-makers or legacy IT/operational processes, Computational and communication costs for large-scale deployment are not fully quantified

Claims (11)

Claim	Direction	Confidence	Outcome	Details
HGA-MADDPG introduces a hierarchical graph attention mechanism to dynamically represent the state of the supply chain network topology. Other	positive	high	dynamic representation of supply chain network topology (method implementation)	0.09
The algorithm quantifies the contribution of individual actions to sub-chain objectives and system-level indicators through local and global credit networks. Other	positive	high	quantification of action contributions to sub-chain and system-level objectives (method implementation)	0.09
An adaptive fusion weight based on marginal returns is designed to dynamically balance local and global credit. Other	positive	high	adaptive weighting between local and global credit (method implementation)	0.09
The paper constructs an adversarial disturbance and resilient training architecture that models three types of disturbances (demand mutation, node failure, transportation delay), adversarial agent injection, a dynamic environment replay buffer, and a two-stage training strategy. Other	positive	high	presence and implementation of adversarial/resilient training components (method implementation)	0.09
In a baseline scenario (four-level supply chain, dynamic environment driven by real data from SCDL and WSN) and compared with eight baseline algorithms, HGA-MADDPG achieves a total cost reduction rate of 26.2%. Organizational Efficiency	positive	high	total cost (operational cost) of the supply chain	26.2% reduction 0.18
In the same baseline scenario, HGA-MADDPG achieves a service level improvement rate of 42.8% compared with eight baseline algorithms. Organizational Efficiency	positive	high	service level (supply/delivery performance)	42.8% improvement 0.18
In the same baseline scenario, HGA-MADDPG controls the stockout rate at 3.2%. Organizational Efficiency	positive	high	stockout rate	3.2% stockout rate 0.18
In an extreme scenario of triple perturbation, HGA-MADDPG achieves a cost deviation rate of 29.6%, which is significantly better than existing methods. Organizational Efficiency	positive	high	cost deviation rate under perturbation	29.6% cost deviation rate 0.18
In the same extreme scenario of triple perturbation, HGA-MADDPG achieves a recovery time of 58 hours, outperforming existing methods. Organizational Efficiency	positive	high	recovery time (hours) after perturbation	58 hours 0.18
HGA-MADDPG maintains a cost reduction rate of 21.5% in a 120-node ultra-large-scale supply chain. Organizational Efficiency	positive	high	total cost (operational cost) in a 120-node supply chain	n=120 21.5% reduction 0.18
Ablation experiments and scalability analysis verify the effectiveness of each core module of HGA-MADDPG. Organizational Efficiency	positive	high	contribution of core modules to algorithm performance (ablation results)	0.18

A hybrid local–global credit multi-agent RL cuts simulated supply-chain costs by about 26% and boosts service levels by roughly 43%, while retaining robustness under multiple simultaneous disruptions and scaling to 120 nodes.