Robust Collaborative Learning by Multi-Agents

In this paper, we introduce a collaborative learning problem that is applicable in multi-agent data mining using heterogeneous computing resources in environments with limited control, resource failures, and communication bottlenecks. Specifically, we consider the scenario in which multiple agents collect noisy and overlapping information regarding an entity, such as a network attribute, which might correspond to multiple models. The agents are unable to share the entire information due to communication bottlenecks and other strategic issues; instead, the agents share their "local estimate" about the entity. The objective is to obtain the best estimate of the true value of the entity based on the local estimates shared by the agents. First, we derive a centralized solution where the locally processed information from each agent is assumed available at a central node. Then, we develop a distributed solution to the problem that is suitable to environments with limited control, resource failures, and communication bottlenecks.