Towards collective wisdom : a novel approach for cross-distribution collaborative learning

Abstract

Collective Intelligence, harnessing the knowledge and skills of individuals, enables emergent behaviours from their interactions, solving complex tasks that would be insurmountable for individuals alone. Integrating AI models, including machine learning, deep learning, and reinforcement learning, as intelligent components within these systems is a logical progression.Despite emerging approaches in collaborative learning, existing literature falls short of addressing all the requisites essential for achieving true collective intelligence in distributed contexts. Specifically, no current method comprehensively ensures model architecture agnosticism, cross-distribution robustness, applicability to fully distributed systems, and real-time operation.This thesis presents a novel approach to collective learning designed to enhance emergent behaviours through collaborative training across heterogeneous environments independent of data distributions and model architectures. By exchanging insights and data subsets, individuals can build a more comprehensive knowledge base by integrating the experiences and perspectives of others into their datasets. This enhances individual and collective performance, enabling agents subject to different environmental conditions to collaborate towards a common goal.The proposed protocol is validated through two case scenarios, testing the protocol with different parameters and logic for data selection, integration, and memory management. These scenarios demonstrate the system’s robustness against data drift and its capacity to foster valuable emergent behaviours under a correctly tuned protocol.Experimental results confirm the protocol’s effectiveness in improving generalisation and performance at individual and group levels. Agents operating as a collective under this protocol achieved system goals in fewer steps and approached optimal solutions, irrespective of heterogeneity in sub-environments and employed architectures.