Gravar-mail: Generalization of value in reinforcement learning by humans