Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a policy neural network used to select actions to be performed by a reinforcement learning agent interacting with an environment. In one aspect, a method includes obtaining path data defining a path through the environment traversed by the agent. A consistency error is determined for the path from a combined reward, first and last soft-max state values, and a path likelihood. A value update for the current values of the policy neural network parameters is determined from at least the consistency error. The value update is used to adjust the current values of the policy neural network parameters.