Combining Reinforcement Learning and Heuristic Optimization: A Model Based on a Deep Q-Network and Graph Neural Networks for Graph Coloring

We present a hybrid approach combining Reinforcement Learning (RL) with the TabuCol, which is a version of tabu search specifically designed for the Graph Coloring Problem (GCP), enhanced by Graph Neural Networks (GNNs), to tackle the GCP. Our model, referred to as RLTB, integrates a Deep Q-Network architecture with multiple GNN layers to select promising (node, color) pairs, guiding the RL agent towards optimal solutions. Experimental evaluations on standard DIMACS graph instances demonstrate that RLTB significantly outperforms existing heuristic algorithms such as Tabucol and a greedy algorithm, surpasses GNN methods from other papers, and shows competitive performance against other outstanding hybrid methods.