UCLA

Exponential growth in the need for low latency offloading of computation was answered by the introduction of edge networks. Since these networks are essentially isolated islands of computing, current prevalent centralized approaches to training learning agents should be adapted to account for the decentralized nature of this new network structure. Since these networks are designed for low latency, cost-awareness must be built into machine learning models when dealing with data streams. Additionally, in order to debias or expand on the locally available data while maintaining edge benefits, multi-agent systems should be constructed to allow for limited coordination outside of a local node.To address these issues, we suggest a novel end-to-end solution that supports the lifetime of a learning agent on the network. We reevaluate how learning agents receive information on an edge network and explore ways for them to communicate and coordinate with other agents efficiently while maintaining context. This thesis will dive into cost-awareness as it pertains to data acquired sequentially and messages exchanged on a network. Additionally, we will showcase our solution for knowledge transfer between remote agents that preserves all the benefits of running in a decentralized network environment.