Conceitos Básicos
By leveraging information sharing between neighboring agents in a communication network, the proposed MT-CO2OL algorithm can achieve improved regret bounds compared to independent learning, especially when neighboring agents operate on similar tasks.
Resumo
The paper introduces MT-CO2OL, a decentralized algorithm for multitask online learning on arbitrary communication networks. The key idea is that agents can fetch predictions from their neighbors and use this information to improve their own learning.
The main highlights and insights are:
MT-CO2OL uses a base algorithm (MT-FTRL) that can operate on a virtual clique of each agent's neighbors. This allows the regret analysis to leverage the performance of MT-FTRL on these local cliques.
The regret bounds of MT-CO2OL depend on the interplay between the task similarities within each agent's neighborhood and the structure of the communication graph. Specifically, the regret scales with the local task variances rather than the global task variance.
For adversarial agent activations, MT-CO2OL achieves regret bounds that improve upon the naive approach of running independent instances of the base algorithm, especially when neighboring agents have similar tasks.
For stochastic agent activations, MT-CO2OL can further improve the regret bounds, matching the best known results for single-task settings.
The authors also provide lower bounds showing the tightness of their results for regular communication graphs.
A differentially private variant of MT-CO2OL is introduced, where the additional regret due to privacy has a mild dependence on the time horizon. This allows identifying privacy thresholds above which sharing information becomes detrimental.
Overall, the paper presents a comprehensive analysis of decentralized multitask online learning, providing both upper and lower regret bounds that highlight the benefits of leveraging neighborhood information.