Learning action models with indeterminate effects