Shepherd: Reinforcement Learning as a Service withDistributed Execution

Most current implementations of Reinforcement Learning agents con-sider that one agent interacts with one environment, and that the agent and envi-ronment run on the same machines. Previous work, such as RL-Glue1, went a stepin the direction of allowing the agent and environment to be different processeson a computer, but a wider separation of the agent and environment is much lesscommon. In this demonstration, we illustrate how Shepherd, a web-service thatallows clients to remotely query a Reinforcement Learning agent for actions, al-lowsmultiple peopleto interact at the same time with asingle agent, on theirphone, over the Internet, without having to install anything. Shepherd ensuresthat knowledge obtained from one client (one person in this demonstration) isquickly leveraged to improve the performance of the agent for the other clients.