Résumé: As the extension of cyber-physical systems (CPSs), cyber-physical-social systems (CPSSs) seamlessly integrate cyber space, physical space, and social space. CPSS provide a more comprehensive smart space of perception for Reinforcement Learning (RL), so as to lead a revolution in Artificial Intelligence (AI), which urgently require innovation in computing system architecture. This paper aims to provide a comprehensive review and perspectives of RL architecture for collaborative computing system in CPSS. Firstly, we analyze CPSS AI's features from data perspective, including multi-modal data fusion from multi-spaces, and rule discover and representation, collaborative computing system architectures for multi-spaces, intelligent decision-making (policy) discovery. After that, we propose to use the action aware transition tensor to fuse the CPSS data for collaborative RL. Then the typical architectures and methods of RL are surveyed. Furthermore, a tensor based unified collaborative computing reinforcement architecture are proposed for CPSS AI, including architecture and optimal policy solution. In the end, we summarize the paper and discuss the future work.

Langue: Anglais