Probabilistic model checking is a technique employed for verifying the correctness of computer systems that exhibit probabilistic behaviour. A related technique is controller synthesis, which generates controllers that guarantee the correct behaviour of the system. Not all controllers can be generated offline, as the relevant information may only be available when the system is running, for example, the reliability of services may vary over time. In this thesis, we propose a framework based on controller synthesis for stochastic games at runtime. We model systems using stochastic two-player games parameterised with data obtained from monitoring of the running system. One player represents the controllable actions of the system, while the other player represents the hostile uncontrollable environment. The goal is to synthesize, for a given property specification, a controller for the first player that wins against all possible actions of the environment player. Initially, controller synthesis is invoked for the parameterised model and the resulting controller is applied to the running system. The process is repeated at runtime when changes in the monitored parameters are detected, whereby a new controller is generated and applied. To ensure the practicality of the framework, we focus on its three important aspects: performance, robustness, and scalability. We propose an incremental model construction technique to improve performance of runtime synthesis. In many cases, changes in monitored parameters are small and models built for consecutive parameter values are similar. We exploit this and incrementally build a model for the updated parameters reusing the previous model, effectively saving time. To address robustness, we develop a technique called permissive controller synthesis. Permissive controllers generalise the classical controllers by allowing the system to choose from a set of actions instead of just one. By using a permissive controller, a computer system can quickly adapt to a situation where an action becomes temporarily unavailable while still satisfying the property of interest. We tackle the scalability of controller synthesis with a learning-based approach. We develop a technique based on real-time dynamic programming which, by generating random trajectories through a model, synthesises an approximately optimal controller. We guide the generation using heuristics and can guarantee that, even in the cases where we only explore a small part of the model, we still obtain a correct controller. We develop a full implementation of these techniques and evaluate it on a large set of case studies from the PRISM benchmark suite, demonstrating significant performance gains in most cases. We also illustrate the working of the framework on a new case study of an open-source stock monitoring application.