Dynamic flexible job shop scheduling (DFJSS), as an extension of job shop scheduling (JSS), has attracted a lot of attention both from academics and industries. The main challenge of DFJSS is to decide which machine to select when routing decision points arrive and which operation to select when sequencing decision points arrive, as the new jobs arrive over time. Among the existing methods for DFJSS, scheduling heuristics are effective methods widely used because of the advantages that are simple to implement and can react in real-time. However, the design of scheduling heuristics requires a lot of domain knowledge and is time-consuming. Genetic programming (GP), as a hyper-heuristic method, has been widely used and achieved great success to evolve scheduling heuristics. The overall goal of this thesis is to improve the effectiveness and efficiency of GP to evolve promising and interpretable scheduling heuristics for solving DFJSS problems. This will be achieved by improving the population diversity for GP and incorporating ensemble, semantic and multi-objective technologies with different strategies.