Abstract
Today, distributed learning and networking are closely intertwined in many applications. On one hand, mobile users need to learn the latest useful information in daily networking systems (e.g., transportation networks and service systems with queueing networks) for better decisions. On the other hand, the networking structure facilitates the use of distributed computing resources to parallelize and process machine learning tasks across multiple nodes or experts, leading to improved model performance and faster convergence. Therefore, this thesis focuses on these two perspectives to do online optimization and mechanism design for distributed learning networks. First, we examine learning for networking, which fundamentally extend traditional networking systems by incorporating the distributed learning of mobile users. Then, we shift our focus to networking for learning, which leverages distributed networking structures to solve machine learning problems.
In the first part of this thesis, we design novel mechanisms, both monetary and non-monetary, to regulate the online distributed learning of selfish users in transportation networks and queueing service systems, leveraging mobile crowdsourcing. First, we consider online pricing for a crowdsourcing platform (e.g., Waze) to economically reward drivers for diverse routing and controlling the freshness of information across all paths. We propose a new backward-clustered computation method and develop an approximation algorithm to alternate different paths to price over time, approaching optimal performance as the number of paths increases. Second, we focus on incentivizing selfish users to learn diverse stochastic paths and achieve the best exploitation-exploration trade-off in congestion games. We consider three scenarios with different types of users: a single myopic user arriving at each time, multiple myopic users arriving at each time, and repeated arrivals of non-myopic non-atomic users. In each scenario, we demonstrate that users’ selfish routing policies result in significant efficiency loss compared to the socially optimal policy that minimizes the long-term cost for all users. To mitigate this efficiency loss, we propose a novel informational (non-monetary) mechanism in each case to regulate selfish users and achieve near-optimal system performance. Finally, we analyze distributed learning in queueing service systems. We show that selfish users tend to over-explore servers, leading to congestion and an infinite price of anarchy (PoA). To address this, we propose a dynamic side-payment mechanism that periodically charges some customers and rewards others to curb over-exploration while maintaining an ex-post budget balance, thereby reducing the PoA to less than 2.
In the second part of this thesis, we perform online optimization and mechanism design to enhance machine learning performance using distributed network structures, focusing on multi-armed bandit (MAB) problems, mixture-of-experts (MoE) models, and continual learning (CL). For MAB problems, we examine competitive non-myopic players with different private beliefs about various arms, where neighboring players choosing the same arm cause conflicts. For example, in Cognitive Radio networks, neighboring nodes using the same channel to transmit data lead to collisions and make their transmission unsuccessful. We consider two scenarios: players with and without prior knowledge of the network topology of their conflict graphs. In both cases, we demonstrate that competition among selfish players results in longer convergence times compared to the socially optimal policy and can lead to arbitrarily large efficiency losses. To address this, we propose novel mechanisms that regulate players’ selfish arm decisions, ensuring truthful reporting and achieving or approaching the social optimum with fast convergence times. We then study how the distributed MoE structure improves the performance of CL compared to a single expert. We theoretically analyze the benefits of MoE by proving that it can diversify its experts to specialize in different tasks, while its router learns to select the appropriate expert for each task. Additionally, we provide explicit expressions for expected catastrophic forgetting and overall generalization error, characterizing the benefits of MoE in CL performance. Interestingly, while adding more experts to the network requires additional rounds before convergence, it may not always enhance learning performance.