The application of reinforcement learning techniques to optimization problems has gained increasing attention due to their adaptability, generalization potential, and capacity to handle complex decision-making processes. This study explores the opportunities and limitations of Q-learning, in the context of the classical Assignment Problem, which plays an important role in transportation logistics and resource allocation scenarios. Four variants of the algorithm were developed and evaluated: a basic version, a version incorporating min-max normalization of cost values, a long-term profitability strategy, and a backward optimization approach. For each of the algorithms, the hyperparameters were optimized using the Optuna library and tests were performed on randomly generated cost matrices of varying dimensions (5, 10, 50, 100, and 200). The quality of the solutions was evaluated based on degradation relative to the optimal objective function value. The time to generate solutions was also measured. The results indicate significant differences in the capabilities of different algorithm variants. The basic Q-learning version is characterized by limited effectiveness and high variability, particularly for larger problem instances. Normalization improved computational efficiency and reduced variance, but did not lead to substantial improvements in solution quality for more complex cases. In contrast, the long-term profitability variant demonstrated notable improvements in both solution quality and stability, especially for smaller and medium-sized problems. The backward optimization variant yielded the highest overall solution quality.
Misztal et al. (Tue,) studied this question.