Empowering Large Language Model Agents through Action Learning