© 2020 Elsevier B.V. Wireless communication and computation technologies are becoming increasingly complex and dynamic due to the sophisticated and ubiquitous Internet of things (IoT) applications. Therefore, future wireless networks and computation solutions must be able to handle these challenges and dynamic user requirements for the success of IoT systems. Recently, learning strategies (particularly deep learning and reinforcement learning) are explored immensely to deal with the complexity and dynamic nature of communication and computation technologies for IoT systems, mainly because of their power to predict and efficient data analysis. Learning strategies can significantly enhance the performance of IoT systems at different stages, including at IoT node level, local communication, long-range communication, edge gateway, cloud platform, and corporate data centers. This paper presents a comprehensive overview of learning strategies for IoT systems. We categorize learning paradigms for communication and computing technologies in IoT systems into reinforcement learning, Bayesian algorithms, stochastic learning, and miscellaneous. We then present research in IoT with the integration of learning strategies from the optimization perspective where the optimization objectives are categorized into maximization and minimization along with corresponding applications. Learning strategies are discussed to illustrate how these strategies can enhance the performance of IoT applications. We also identify the key performance indicators (KPIs) used to evaluate the performance of IoT systems and discuss learning algorithms for these KPIs. Lastly, we provide future research directions to further enhance IoT systems using learning strategies