Model-and-indicator-based GRASP-VNS for Two Problems concerning Intensity Modulated Radiotherapy Planning
Cancer, Radiotherapy, GRASP, Variable Neighborhood Search, Mathematical Programming, Artificial Intelligence, Automata Learning.
Intensity-modulated radiotherapy (IMRT) is a widely used cancer treatment. Planning this type of treatment involves two complex computational problems related to the choice of beam angles to irradiate the patient and the intensity that each beam must have so that cancer cells are killed, and at the same time avoid reaching regions with healthy tissue. Metaheuristics have been widely used to address complex problems. Hybridization of metaheuristics often results in methods that are even more effective than metaheuristics used alone. In the context of hybridization, there are also matheuristics, which are unions of metaheuristics with mathematical programming. In this context, the research reported in this work has been added. An algorithm is proposed that hybridizes the GRASP (Greedy Random Adaptive Search Procedure) and VNS (Variable Neighborhood Search) meta-heuristics with mathematical programming models to address the two problems mentioned above. A third approach based on automaton learning, called GRASP-VNS-IA, was also explored to determine the execution order of VNS neighborhoods. Of the four models used, two were proposed in this study. The solutions produced by the algorithm are evaluated using an indicator that combines four indicators, three of which are proposed in this study. GRASP-VNS was compared with GRASP and GRASP-VNS-IA. The algorithms were tested on a set of ten liver cancer instances that are known to be challenging. The results produced by the algorithms were evaluated using quality indicators and histograms. Statistical tests were used to support the conclusions regarding the behavior of the algorithms.