Generating and updating rescheduling knowledge that can be used in real time has become a key issue in reactive scheduling due to the dynamic and uncertain nature of industrial environments and the emergent trend towards cognitive systems in production planning and execution control. Disruptive events have a significant impact on the feasibility of plans and schedules. In this work, the automatic generation and update through learning of rescheduling knowledge using simulated transitions of abstract schedule states is proposed. An industrial example where a current schedule must be repaired in response to unplanned events such as the arrival of a rush order, raw material delay, or an equipment failure which gives rise to the need for rescheduling is discussed. A software prototype (SmartGantt) for interactive schedule repair in real-time is presented. Results demonstrate that responsiveness is dramatically improved by using relational reinforcement learning and relational abstractions to develop a repair policy. (C) 2012 Elsevier Ltd. All rights reserved.