PPO-STGNN: A Proximal Policy Optimization Approach with Spatio-Temporal Graph Neural Networks for DAG Task Scheduling in Cloud-Edge-End Computing
Yangshuo Qi, Chenwei Wang, Zihan Shen, Songlin Sun
cs.AI
Quick verdict
claims load-balance wins with PPO+STGNN for cloud-edge-end DAG scheduling, but reports the worst SLR of all baselines and internally inconsistent numbers.
Summary
The paper proposes PPO-STGNN, a deep reinforcement learning scheduler for DAG-structured workloads in heterogeneous cloud-edge-end environments. An STGNN encoder combines a graph attention network over the physical resource graph with a GRU over its temporal evolution, while a GCN encodes the DAG task topology; the concatenated representation feeds a PPO actor-critic that selects joint task-node actions with topological masking, trained with a dense-plus-terminal reward and behavior-cloning pretraining on HEFT-generated expert trajectories. On a single synthetic scenario, PPO-STGNN achieves the best makespan (124.8 s) and load balance (127.9, 21.3% below HEFT) among FCFS, LeastLoad, HEFT, and Greedy, but the worst schedule length ratio (11.6 vs HEFT's 7.8). Against its own ablations the authors report faster convergence with validation makespan below 60 s, versus roughly 110 s for MLP-PPO. Major limitations: no code, simulator, workload sizes, or hyperparameters; single-workload evaluation without repeats or statistical tests; unexplained contradictions between Table 1 (124.8 s, SLR 11.6) and Section 4.3 curves (<60 s, SLR 15-20); LB values inconsistent with the stated variance-of-normalized-loads definition; and no external DRL baselines.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 16 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- PPO-STGNN achieves the best makespan (124.8 s) among FCFS, LeastLoad, HEFT, and Greedy on the tested cloud-edge-end DAG workload (Table 1).
- PPO-STGNN substantially improves load balancing, reducing LB to 127.9 - 21.3% below HEFT's 162.5 - while maintaining low completion time.
- Spatio-temporal graph encoding outperforms MLP and static-GNN encoders in the same PPO framework, converging faster with validation makespan below 60 s (Sec. 4.3).
- Multi-teacher behavior cloning on HEFT-generated expert trajectories mitigates the cold-start problem of DRL in large joint action spaces (claimed, not ablated).
Strengths
Joint spatio-temporal and topological encoding is a sensible architecture
Sec. 3.1 pairs a latency/bandwidth-aware GAT (Eq. 3) plus a GRU over a history window (Eq. 4) for the resource graph with a GCN (Eq. 5) over DAG features (depth, in/out-degree, critical-path score), directly addressing the static-topology limitation the authors attribute to prior GRL schedulers.
Appropriate action-space design for dependency constraints
The joint task-node action a = Index(v) x K + Index(m) (Eq. 6) with a topological action-mask mechanism (Sec. 3.2) cleanly enforces precedence feasibility, and the dense-plus-terminal reward separates step costs from makespan/SLR credit.
BC pretraining against heuristic experts is a credible cold-start remedy
Comparatively honest reporting of the SLR loss
Weaknesses
Mutually contradictory numbers between the two result sections
Worst SLR of all five methods despite SLR being a claimed objective
LB values irreconcilable with the metric definition
No reproducibility substrate
No external DRL baselines and an unevaluated energy dimension
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
PPO-STGNN (arXiv 2609.03503) targets DAG task scheduling across hierarchical cloud-edge-end infrastructures. The claimed contributions are: (1) a heterogeneous spatio-temporal scheduling model spanning cloud/edge/end node classes with link constraints; (2) an STGNN+PPO architecture in which a graph attention network (GAT) encodes the physical resource graph, a GRU encodes its temporal evolution, and a GCN encodes the DAG topology; (3) a multi-teacher behavior-cloning pretraining stage using heuristic expert trajectories. The headline empirical claims are that PPO-STGNN attains the best makespan (124.8 s, Table 1) and the best load balance (127.9, 21.3% below HEFT's 162.5) against FCFS, LeastLoad, HEFT, and Greedy, and that spatio-temporal encoding beats MLP and static-GNN encoders within the same PPO framework (Sec. 4.3). Notably, the abstract claims joint minimization of makespan *and* SLR, yet the reported SLR (11.6) is the worst of all five methods.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PPO-STGNN?
PPO-STGNN is a deep reinforcement learning scheduler proposed for DAG-structured task graphs in cloud-edge-end computing environments. It encodes the physical resource graph with a graph attention network plus a GRU over time, encodes the DAG task topology with a GCN, and trains a PPO agent (warm-started by behavior cloning on HEFT-generated schedules) to assign tasks to nodes, targeting makespan, schedule length ratio, and load balance.
How does PPO-STGNN compare with HEFT?
In the paper's single reported experiment, PPO-STGNN achieves a lower makespan (124.8 s vs 133.7 s) and better load balancing (127.9 vs 162.5), but a worse schedule length ratio (11.6 vs HEFT's best-of-all-methods 7.8). The comparison is on one synthetic workload with no variance reported.
Is code or data available for PPO-STGNN?
No. The paper names no simulator, gives no node counts, DAG sizes, or training hyperparameters, and provides no code or data release, so the reported numbers cannot currently be reproduced independently.
Why does the paper report two different makespan values for PPO-STGNN?
Table 1 reports 124.8 s for the baseline comparison (FCFS, LeastLoad, HEFT, Greedy), while Section 4.3 describes validation-training convergence below 60 s against its own RL ablations (MLP-PPO at ~110 s, PPO-StaticGNN fluctuating). The paper does not explain the discrepancy; this is one of the review's central open questions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure contents (Figs. 1-4) are absent from the extracted text; only captions and the paper's own verbal descriptions of Fig. 3/4 trends are available, so no independent reading of the curves was possible. The provided text itself is complete (ends cleanly at reference [15]).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.