Search for: All records

Mobile devices such as smartphones, laptops, and tablets can often connect to multiple access networks (e.g., Wi-Fi, LTE, and 5G) simultaneously. Recent advancements facilitate seamless integration of these connections below the transport layer, enhancing the experience for apps that lack inherent multi-path support. This optimization hinges on dynamically determining the traffic distribution across networks for each device, a process referred to as \textit{multi-access traffic splitting}. This paper introduces \textit{NetworkGym}, a high-fidelity network environment simulator that facilitates generating multiple network traffic flows and multi-access traffic splitting. This simulator facilitates training and evaluating different RL-based solutions for the multi-access traffic splitting problem. Our initial explorations demonstrate that the majority of existing state-of-the-art offline RL algorithms (e.g. CQL) fail to outperform certain hand-crafted heuristic policies on average. This illustrates the urgent need to evaluate offline RL algorithms against a broader range of benchmarks, rather than relying solely on popular ones such as D4RL. We also propose an extension to the TD3+BC algorithm, named Pessimistic TD3 (PTD3), and demonstrate that it outperforms many state-of-the-art offline RL algorithms. PTD3's behavioral constraint mechanism, which relies on value-function pessimism, is theoretically motivated and relatively simple to implement. 

The problem of two-player zero-sum Markov games has recently attracted increasing interests in theoretical studies of multi-agent reinforcement learning (RL). In particular, for finite-horizon episodic Markov decision processes (MDPs), it has been shown that model-based algorithms can find an ϵ-optimal Nash Equilibrium (NE) with the sample complexity of O(H3SAB/ϵ2), which is optimal in the dependence of the horizon H and the number of states S (where A and B denote the number of actions of the two players, respectively). However, none of the existing model-free algorithms can achieve such an optimality. In this work, we propose a model-free stage-based Q-learning algorithm and show that it achieves the same sample complexity as the best model-based algorithm, and hence for the first time demonstrate that model-free algorithms can enjoy the same optimality in the H dependence as model-based algorithms. The main improvement of the dependency on H arises by leveraging the popular variance reduction technique based on the reference-advantage decomposition previously used only for single-agent RL. However, such a technique relies on a critical monotonicity property of the value function, which does not hold in Markov games due to the update of the policy via the coarse correlated equilibrium (CCE) oracle. Thus, to extend such a technique to Markov games, our algorithm features a key novel design of updating the reference value functions as the pair of optimistic and pessimistic value functions whose value difference is the smallest in the history in order to achieve the desired improvement in the sample efficiency. 

Network telemetry systems have become hybrid combinations of state-of-the-art stream processors and modern programmable data-plane devices. However, the existing designs of such systems have not focused on ensuring that these systems are also deployable in practice, i.e., able to scale and deal with the dynamics in real-world traffic and query workloads. Unfortunately, efforts to scale these hybrid systems are hampered by severe constraints on available compute resources in the data plane (e.g., memory, ALUs). Similarly, the limited runtime programmability of existing hardware data-plane targets critically affects efforts to make these systems robust. This paper presents the design and implementation of DynaMap, a new hybrid telemetry system that is both robust and scalable. By planning for telemetry queries dynamically, DynaMap allows the remapping of stateful dataflow operators to data-plane registers at runtime. We model the problem of mapping dataflow operators to data-plane targets formally and develop a new heuristic algorithm for solving this problem. We implement our algorithm in prototype and demonstrate its feasibility with existing hardware targets based on Intel Tofino. Using traffic workloads from different real-world production networks, we show that our prototype of DynaMap improves performance on average by 1-2 orders of magnitude over state-of-the-art hybrid systems that use only static query planning. 

Recent studies in reinforcement learning (RL) have made significant progress by leveraging function approximation to alleviate the sample complexity hurdle for better performance. Despite the success, existing provably efficient algorithms typically rely on the accessibility of immediate feedback upon taking actions. The failure to account for the impact of delay in observations can significantly degrade the performance of real-world systems due to the regret blow-up. In this work, we tackle the challenge of delayed feedback in RL with linear function approximation by employing posterior sampling, which has been shown to empirically outperform the popular UCB algorithms in a wide range of regimes. We first introduce Delayed-PSVI, an optimistic value-based algorithm that effectively explores the value function space via noise perturbation with posterior sampling. We provide the first analysis for posterior sampling algorithms with delayed feedback in RL and show our algorithm achieves $$\widetilde{O}(\sqrt{d^3H^3 T} + d^2H^2 E[\tau])$$ worst-case regret in the presence of unknown stochastic delays. Here $$E[\tau]$$ is the expected delay. To further improve its computational efficiency and to expand its applicability in high-dimensional RL problems, we incorporate a gradient-based approximate sampling scheme via Langevin dynamics for Delayed-LPSVI, which maintains the same order-optimal regret guarantee with $$\widetilde{O}(dHK)$$ computational cost. Empirical evaluations are performed to demonstrate the statistical and computational efficacy of our algorithms. 

The remarkable success of the use of machine learning-based solutions for network security problems has been impeded by the developed ML models’ inability to maintain efficacy when used in different network environments exhibiting different network behaviors. This issue is commonly referred to as the generalizability problem of ML models. The community has recognized the critical role that training datasets play in this context and has developed various techniques to improve dataset curation to overcome this problem. Unfortunately, these methods are generally ill-suited or even counterproductive in the network security domain, where they often result in unrealistic or poor-quality datasets. To address this issue, we propose a new closed-loop ML pipeline that leverages explainable ML tools to guide the network data collection in an iterative fashion. To ensure the data’s realism and quality, we require that the new datasets should be endogenously collected in this iterative process, thus advocating for a gradual removal of data-related problems to improve model generalizability. To realize this capability, we develop a data-collection platform, netUnicorn, that takes inspiration from the classic “hourglass” model and is implemented as its “thin waist" to simplify data collection for different learning problems from diverse network environments. The proposed system decouples data-collection intents from the deployment mechanisms and disaggregates these high-level intents into smaller reusable, self-contained tasks. We demonstrate how netUnicorn simplifies collecting data for different learning problems from multiple network environments and how the proposed iterative data collection improves a model’s generalizability 

The application of the latest techniques from artificial intelligence (AI) and machine learning (ML) to improve and automate the decision-making required for solving real-world network security and performance problems (NetAI, for short) has generated great excitement among networking researchers. However, network operators have remained very reluctant when it comes to deploying NetAIbased solutions in their production networks. In Part I of this manifesto, we argue that to gain the operators' trust, researchers will have to pursue a more scientific approach towards NetAI than in the past that endeavors the development of explainable and generalizable learning models. In this paper, we go one step further and posit that this opening up of NetAI research will require that the largely self-assured hubris about NetAI gives way to a healthy dose humility. Rather than continuing to extol the virtues and magic of black-box models that largely obfuscate the critical role of the utilized data play in training these models, concerted research efforts will be needed to design NetAI-driven agents or systems that can be expected to perform well when deployed in production settings and are also required to exhibit strong robustness properties when faced with ambiguous situations and real-world uncertainties. We describe one such effort that is aimed at developing a new ML pipeline for generating trained models that strive to meet these expectations and requirements. 

The application of the latest techniques from artificial intelligence (AI) and machine learning (ML) to improve and automate the decision-making required for solving real-world network security and performance problems (NetAI, for short) has generated great excitement among networking researchers. However, network operators have remained very reluctant when it comes to deploying NetAI-based solutions in their production networks, mainly because the black-box nature of the underlying learning models forces operators to blindly trust these models without having any understanding of how they work, why they work, or when they don't work (and why not). Paraphrasing [1], we argue that to overcome this roadblock and ensure its future success in practice, NetAI has to get past its current stage of explorimentation, or the practice of poking around to see what happens and has to start employing tools of the scientific method. 

We study linear bandits when the underlying reward function is not linear. Existing work relies on a uniform misspecification parameter $$\epsilon$$ that measures the sup-norm error of the best linear approximation. This results in an unavoidable linear regret whenever $$\epsilon > 0$$. We describe a more natural model of misspecification which only requires the approximation error at each input $$x$$ to be proportional to the suboptimality gap at $$x$$. It captures the intuition that, for optimization problems, near-optimal regions should matter more and we can tolerate larger approximation errors in suboptimal regions. Quite surprisingly, we show that the classical LinUCB algorithm — designed for the realizable case — is automatically robust against such gap-adjusted misspecification. It achieves a near-optimal $$\sqrt{T}$$ regret for problems that the best-known regret is almost linear in time horizon $$T$$. Technically, our proof relies on a novel self-bounding argument that bounds the part of the regret due to misspecification by the regret itself.