The ambition of Predictive State Representations (PSRs) is to bypass the need for defining state spaces in reinforcement learning problems. The idea is that predictions of observation sequences conditioned on action sequences can serve as sufficient statistics for choosing (or learning to choose) optimal actions.