Genetic Programming Hyper-Heuristic for Emergency Medical Dispatching

Abstract

Emergency Medical Dispatch (EMD) defines the healthcare task concerning the assignment of paramedic resources (equipment and people) to emergencies in the community. Due to the random and unpredictable nature of such tasks, EMD is highly dynamic. Research exists using Machine Learning to automatically learn heuristics for EMD via the Genetic Programming Hyper Heuristic (GPHH) technique. This project seeks to improve upon such existing research by implementing multi-fidelity techniques. In particular, by training GPHH models on city networks (graphs) of varying fidelity. The models produce lower-fidelity graphs by subdividing a source graph into boxes, and quantising locations within each box to a single location. The fidelity of graphs is then dictated by the size of the subdivided boxes. The computation savings afforded by multi-fidelity techniques decrease model training time on graphs generated from real-world ambulance coverage areas, whilst not significantly impacting training or test performance. The multi-fidelity models were evaluated against pre-existing models on two real-world graphs, the Wellington and Christchurch ambulance coverage areas, to investigate the feasibility of applying multi-fidelity techniques to EMD.