In the highly interdependent environment of a large city, failures in the Electrical Distribution System (EDS) can cause direct or indirect consequences to other critical infrastructures and the well-being of the citizens. To increase the resilience of the supply of electricity to the city, this work combines the pre-training of an AI agent and very fast calculation of the optimum recovery path after the number and location of the electrical faults are known. In the introduced Soft-Hard Optimal Convergence (SHOC) method, machine learning techniques are used to train an AI agent with thousands of off-line scenarios for optimum system restoration. In real-time, after the actual fault information is known, the agent will provide a subset of solutions (soft solution) to be considered for hard optimization algorithms. The Infrastructure Interdependencies Simulator (i2SIM) is used to assist the prioritization of the sequence of fault recovery and topological reconfiguration to minimize the black-out time of the most critical loads. A 70-node distribution system case is used to demonstrate the proposed methodology, with solution times in the order of seconds to find the optimum repair sequence and switches topological reconfiguration to optimize the city's resilience index.