Training more effective learned optimizers