A deep reinforcement learning assisted simulated annealing algorithm for a maintenance planning problem
针对同时考虑劳动力规划、培训和备件库存的联合维护规划问题，提出一种结合深度强化学习与模拟退火的混合算法，在小规模实例上能找到最优解，且优于传统元启发式和机器学习算法。
Abstract Maintenance planning aims to improve the reliability of assets, prevent the occurrence of asset failures, and reduce maintenance costs associated with downtime of assets and maintenance resources (such as spare parts and workforce). Thus, effective maintenance planning is instrumental in ensuring high asset availability with the minimum cost. Nevertheless, to find such optimal planning is a nontrivial task due to the (i) complex and usually nonlinear inter-relationship between different planning decisions (e.g., inventory level and workforce capacity), and (ii) stochastic nature of the system (e.g., random failures of parts installed in assets). To alleviate these challenges, we study a joint maintenance planning problem by considering several decisions simultaneously, including workforce planning, workforce training, and spare parts inventory management. We develop a hybrid solution algorithm ( $$\mathcal {DRLSA}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>DRLSA</mml:mi> </mml:math> ) that is a combination of Double Deep Q-Network based Deep Reinforcement Learning (DRL) and Simulated Annealing (SA) algorithms. In each episode of the proposed algorithm, the best solution found by DRL is delivered to SA to be used as an initial solution, and the best solution of SA is delivered to DRL to be used as the initial state. Different from the traditional SA algorithms where neighborhood structures are selected only randomly, the DRL part of $$\mathcal {DRLSA}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>DRLSA</mml:mi> </mml:math> learns to choose the best neighborhood structure to use based on experience gained from previous episodes. We compare the performance of the proposed solution algorithm with several well-known meta-heuristic algorithms, including Simulated Annealing, Genetic Algorithm (GA), and Variable Neighborhood Search (VNS). Further, we also develop a Machine Learning (ML) algorithm (i.e., K-Median) as another benchmark in which different properties of spare parts (e.g., failure rates, holding costs, and repair rates) are used as clustering features for the ML algorithm. Our study reveals that the $$\mathcal {DRLSA}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>DRLSA</mml:mi> </mml:math> finds the optimal solutions for relatively small-size instances, and it has the potential to outperform traditional meta-heuristic and ML algorithms.