A Framework for Distributed Reinforcement Learning Pan Gu and Antnony B. Maddox 330 Snell Center, Northeastern University, 360 Huntington Avenue, Bonton, MA 02115 pan@splinter.coe.neu.edu amaddox@gse.ucla.edu Abstract This paper proposes a novel learning model, called the distributed reinforcemnet learning model (DRLM), that allows distributed agents to learn multiple interrelated tasks in a real-time environment. DRLM consists of a hidden task model (HTM) used for dealing with incomplete perception, a composite state model (CSM) for interdependency between tasks, and a Q- learing subsystem (QLS) for updating action merit. In this paper, we also present a distributed Q-learning algorithm and an architecture that allows agents to reward their peers' actions and share their experience. DRLM is successfully implemented in a flexible manufactuing system where sensors (modeled as agents) have to learn to communicate with humans about the material handling activities using graphical actions such as displays and animaiton.