Developing proper maintenance and rehabilitation investment plans is vital for prolonging the service life of road infrastructures while preserving the required service level under capital constraints. This paper proposes a reinforcement learning approach for determining an optimal policy of selecting maintenance, repair and rehabilitation alternatives for a network of road infrastructure facilities. The proposed approach is based on a policy gradient method and overcomes the computational complexity of optimisation problems due to a large number of possible combinations of network conditions and maintenance, repair and rehabilitation alternatives. The developed optimal management policy takes into consideration interdependencies among infrastructure facilities in a road network. Numerical studies on concrete bridge decks in road networks are performed to demonstrate the advantage, feasibility and capability of the proposed approach.