The article introduces MT-CO 2OL, a new decentralized algorithm for multitask online learning. This algorithm operates in a setting where agents exchange information via a communication network. The regret, or the measure of prediction error, of this algorithm depends on the interplay between task similarities and network structure. The study shows that MT-CO 2OL’s regret is never worse than when agents do not share information. Moreover, the regret significantly improves when neighboring agents work on similar tasks. The algorithm also maintains privacy with a negligible impact on the regret when the losses are linear. Experimental support for these findings is provided.
Publication date: 27 Oct 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2310.17385
