A New Adaptation Mechanism of the ALNS Algorithm Using Reinforcement Learning

Adaptive Large Neighborhood Search (ALNS) is used to solve NP-hard practical problems. Selecting operators and changing parameters to match a specific purpose is a difficult aspect of metaheuristic design. Our proposal concerns ALNS operator selection. Classical ALNS uses “roulette wheel selection” (RWS) to pick operators during the search phase. Choosing operators with RWS is a big challenge because an operator will almost always take the best spot in the roulette, whereas evolutionary algorithms require a balance between exploration and exploitation. We provide an improved ALNS metaheuristic for the capacitated vehicle routing problem (CVRP) that balances exploration and exploitation. The suggested strategy favors the most successful operators using reinforcement learning, notably the Q-learning algorithm. The experimental study shows that the suggested approach works well and is comparable to the classic ALNS.