PAPER: INTRODUCTIONThe growth of connected device networks, such as the Internet of Things (IoT), has led to a surge in data generation. Traditionally, Cloud Computing handled these computational demands, but increased network traffic and latency became apparent as these networks expanded Min et al. (2019). Edge Computing (EC) extends the cloud by bringing computational resources closer to end-users, addressing latency and traffic issues Varghese & Buyya (2018). Despite distinguishing between Mobile Edge Computing (MEC) and Fog computing, this paper treats them interchangeably, focusing on their goal of minimizing device-to-cloud distances Yu et al. (2020). The EC paradigm distributes computational resources, making centralized network orchestration inefficient. Centralization would require aggregating data at a single node, straining the network, and creating a single point of failure Baek & Kaddoum (2023). This highlights the value of decentralized orchestration, particularly through Task Offloading (TO). Optimal TO in such distributed environments involves managing multiple factors, including task latency, energy consumption, and task completion reliability Zhu et al. (2019). Traditional optimization methods often struggle to efficiently manage these complex systems, due to the dynamic, time-varying, and complex environments of Edge Systems Xu et al. (2018). Reinforcement Learning (RL) Baek & Kaddoum (2023); Zhu et al. (2019), is a powerful candidate and dominant approach to solving the TO problem. Specifically, Multi-Agent Reinforcement Learning (MARL) has been explored as a promising solution for decentralized orchestration in Edge Systems Baek & Kaddoum (2023); Gao et al. (2022). The ability of MARL agents to iteratively learn optimal strategies through simultaneous interaction with an environment makes them particularly suited for decentralized edge systems Lin et al. (2023); Zhang et al. (2023). Due to the nature of MARL, it is common to have some form of message exchange Zhang et al. (2018); Baek & Kaddoum (2023) between participants, as this reduces the uncertainty generated by having multiple agents interacting simultaneously, making it particularly suitable for Federated Learning (FL), which has recently gained academic interest as an efficient and distributed approach to agent cooperation in learning Consul et al. (2024). When FL is applied to MARL and the agents only have partial observability of the state, as is commonly the case in decentralized systems where obtaining information about all nodes comes at a premium, it creates a paradigm known as Vertical Federated Reinforcement Learning (VFRL) Qi Under review as a conference paper at ICLR 2026ICLR et al. (2021)). The MARL problem is transformed from one in which agents focus solely on their own objectives into a global optimization problem that accounts for the collective objectives of the participants in the federation. FL also mitigates the strain on the network by avoiding the exchange of large amounts of information, since agents only need to periodically share their learned updates that are aggregated into a global unified model solving the global objective. This enables agents to benefit from each other's knowledge while minimizing communication overhead. However, conventional FL suffers when stragglers delay aggregation or drop updates, reducing training efficiency and wasting samples. This can be addressed by adopting a buffered semi-asynchronous strategy, in which faster nodes continue contributing updates without waiting, while slower nodes are still able to align with the evolving global critic. FAuNO adopts a buffered semi-asynchronous strategy, where faster nodes continue contributing updates without waiting, while slower nodes are still able to align with the evolving global critic. In this way, we extend Federated Buffering Nguyen et al. (2021) to reinforcement learning, enabling continuous local training without being bottlenecked by stragglers. We summarize the motivations and principal contributions of this work below.
================================================================================

REVIEW
--------------------------------------------------------------------------------
**Summary Of The Paper**

The paper presents **FAuNO**, a novel federated reinforcement learning (FRL) framework tailored for task offloading in edge computing systems. It leverages **multi-agent reinforcement learning (MARL)**, specifically **Proximal Policy Optimization (PPO)**, in conjunction with **buffered semi-asynchronous aggregation** inspired by **FedBuff** to address key challenges in decentralized edge environments—namely, partial observability, heterogeneity, and straggler delays. The task offloading problem is formulated as a **partially observable Markov game (POMG)**, capturing the distributed and uncertain nature of edge systems. FAuNO employs a **federated critic** while keeping **local actors**, thereby achieving a balance between global coordination and decentralized execution. The proposed method is evaluated in simulated edge environments using **PeersimGym** and **Ether**, with extensive ablation studies analyzing performance under varying levels of heterogeneity, aggregation thresholds, and federated component configurations. The results indicate that FAuNO outperforms existing heuristics and FRL baselines in terms of task completion rate and latency.

---

**Strengths**

1. **Novel Integration of Techniques**: FAuNO uniquely combines **FedBuff**, a buffered semi-asynchronous aggregation technique originally proposed for non-MARL settings, with **PPO-based actor-critic MARL** in a federated setting. This integration enables continuous local training despite straggler delays, offering a fresh angle for decentralized edge task offloading.

2. **Comprehensive Evaluation**: The paper conducts extensive ablation studies on **heterogeneous workloads**, **aggregation thresholds**, and **component-level federated designs** (critic-only, actor-only, etc.). These analyses help isolate the contributions of various aspects of the framework and demonstrate its robustness.

3. **Realistic Simulations**: Experiments are conducted using **PeersimGym** and **Ether**, generating both synthetic and realistic topologies. This ensures that the findings reflect conditions relevant to actual edge systems, rather than idealized or overly simplified scenarios.

4. **Clear Problem Formulation**: The paper formally defines the task offloading problem as a **constrained optimization problem** and reformulates it as a **POMG**, providing a solid foundation for applying MARL techniques.

5. **Reproducibility Efforts**: An anonymous repository containing the full codebase, hyperparameters, and training scripts is provided, facilitating replication of the experiments and verification of results.

---

**Weaknesses**

1. **Ambiguous Formalization of POMG (Severe)**: The definition of the POMG lacks clarity regarding the **explicit formalization of partial observability**. While the paper mentions that agents observe only local and neighbor information, it fails to rigorously define the mapping from the global state $ S $ to the observed states $ o_i $, leaving room for interpretation and limiting the generality of the formulation.

2. **Insufficient Differentiation from Prior Work (Major)**: Although the paper cites previous works like **FEDOR** and **FLoadNet**, it does not clearly articulate **what distinguishes FAuNO from these existing federated MARL approaches**. For instance, FEDOR already uses a federated critic and FedAvg, yet FAuNO claims to be "first" in integrating FedBuff with PPO. More comparative analysis is needed to establish novelty.

3. **Overreliance on Simulation (Major)**: All experiments are conducted in **simulated environments** (PeersimGym, Ether). There is no mention of **real-world deployment or validation** on physical edge infrastructure, which limits the practical applicability and generalizability of the findings.

4. **Lack of Statistical Rigor (Moderate)**: While the paper reports averages over 40 episodes, it **does not provide statistical significance checks** (e.g., confidence intervals, t-tests) to determine whether the improvements over baselines are statistically meaningful.

5. **Ambiguity in Critic Agreement Metric (Moderate)**: The **global disagreement score** (equations 12–13) normalizes critic outputs by **task arrival rate** $ \lambda_i $, which may obscure true differences in policy behavior. The rationale behind this choice is unclear, and it raises concerns about the validity of the metric in assessing policy consistency across heterogeneous settings.

---

**Questions For The Authors**

1. In Section 3, the POMG is defined with a global state $ S $, but the mapping from $ S $ to each agent's observation $ o_i $ is not explicitly described. Can you clarify how the **partial observability assumption** is formalized, and what constitutes the **observation functions $ O_i $**?

2. The paper claims FAuNO is the **first** to integrate FedBuff with PPO in a federated MARL context. However, **FEDOR** (Zang et al., 2022) already implements federated critics and FedAvg. Please clarify **how FAuNO differs from FEDOR and other similar frameworks** in terms of architecture, aggregation mechanism, and performance characteristics.

3. The **normalization of critic outputs by $ \lambda_i $** in the critic agreement protocol seems unconventional. Why is this done, and how does it affect the **interpretation of policy divergence** across heterogeneous agents?

4. In Equation 11, the global critic update is expressed as a **weighted average of gradients**. How is the **buffer size $ K $** determined dynamically, and what guarantees exist for **convergence under asynchronous updates**?

5. The paper evaluates FAuNO using **PeersimGym and Ether**, but no **comparison to real-world edge networks** is included. Have the authors tested FAuNO on real hardware or deployed it in a controlled edge environment?

---

**Limitations Not Addressed By The Authors**

1. **No Analysis of Energy Consumption**: The paper acknowledges energy costs as an omission in the introduction but offers **no analysis of how FAuNO’s offloading decisions affect energy consumption**. This is a critical concern for edge computing, where energy efficiency is paramount.

2. **Adversarial Behavior Ignored**: The paper assumes all nodes are honest and cooperative. No consideration is given to **adversarial or Byzantine behavior**, which is a crucial challenge in decentralized systems. How would FAuNO respond to **malicious updates** or **node failures**?

3. **Scalability Concerns**: The use of a **centralized global manager (GM)** introduces a **single point of failure and potential bottleneck** in large-scale deployments. The paper does not explore **decentralized or hierarchical critic architectures** to alleviate this issue.

4. **Limited Baseline Comparisons**: The results are compared only to **LQ and SCOF**, but not to **other prominent FRL methods** such as FedDRL or FEDOR. This limits the strength of the claim that FAuNO “outperforms or matches FRL and heuristic baselines.”

5. **Missing Real-World Validation**: The absence of **any real-world deployment or benchmarking** weakens the argument for FAuNO’s practical relevance and effectiveness in real edge environments.

---

**Soundness**: 3 (Good)

While the paper presents a coherent technical proposal supported by simulations, certain aspects of the methodology (e.g., POMG formalism, critic agreement metric) lack sufficient detail and rigor, and the assumptions (e.g., honest nodes, no energy considerations) raise doubts about broader applicability.

**Contribution**: 3 (Good)

The paper contributes a novel application of FedBuff with PPO in a federated MARL setting for edge offloading, along with comprehensive ablations. However, the novelty is somewhat diluted by insufficient differentiation from prior work.

**Confidence**: 4 (High)

The paper is well-written, and the experiments appear sufficiently detailed. The inclusion of an open-source repository enhances confidence in replicability.

**Rating**: 7 (Accept)

The paper makes a valuable contribution to the field of federated MARL for edge computing. Its methodology is sound and well-motivated, though limitations in real-world validation and depth of novelty reduce its impact slightly. With revisions addressing the above concerns, it merits acceptance.

**Brief Justification For Rating**: The paper presents a novel and well-executed approach to federated MARL for edge task offloading, with strong empirical evaluation and good reproducibility efforts. However, the lack of real-world validation, ambiguous definitions, and insufficient differentiation from prior work limit its overall impact. With appropriate revisions, it deserves acceptance.

================================================================================
