Multipath TCP is widely adopted to enhance connection quality-of-service by leveraging multiple network pathways on modern devices. However, the evolution of its core congestion control is hindered by the OS kernel, whose monolithic design imposes high development overhead and lacks the resource flexibility required for data-driven methods. Furthermore, inherent noise in network statistics induces a partial observability problem, which can mislead data-driven methods like Deep Reinforcement Learning. To bridge this gap, we propose Jazz, a system that re-architects multipath congestion control through a decoupled architecture that separates the decision-making ``brain'' from the in-kernel datapath, enabling it to operate on an external (edge) entity. At its core, Jazz employs a Transformer-based agent that processes sequences of historical observations to overcome the partial observability of single-step reinforcement learning. This allows it to learn and master fluctuating link conditions and intricate cross-path dependencies. Tested on a dual-band (5GHz/6GHz) Wi-Fi testbed, our implementation improves bandwidth efficiency by at least 2.85\% over conventional methods and maintains 96.2\% performance under 1\% packet loss, validating this design as a practical blueprint for agile network intelligence.

翻译：多路径TCP被广泛采用，通过利用现代设备上的多个网络路径来提升连接服务质量。然而，其核心拥塞控制的发展受到操作系统内核的阻碍，其整体式设计带来了高昂的开发开销，并且缺乏数据驱动方法所需的资源灵活性。此外，网络统计信息中固有的噪声会引发部分可观测性问题，这可能误导深度强化学习等数据驱动方法。为弥补这一差距，我们提出了Jazz系统，该系统通过解耦架构重新设计了多路径拥塞控制，将决策“大脑”与内核内数据路径分离，使其能够在外部（边缘）实体上运行。Jazz的核心采用了一个基于Transformer的代理，该代理处理历史观测序列，以克服单步强化学习的部分可观测性限制。这使得它能够学习并掌握波动的链路条件以及复杂的跨路径依赖关系。在双频（5GHz/6GHz）Wi-Fi测试平台上进行的测试表明，我们的实现相比传统方法将带宽效率提升了至少2.85%，并在1%丢包率下保持了96.2%的性能，验证了该设计作为敏捷网络智能实用蓝图的可行性。