CIM Bulletin, Vol. 3, No. 8, 2008
H. Askari-Nasab and K. Awuah-Offei
Long-term production scheduling and sequencing are the critical factors in gaining the economic objectives in surface mining. Determination of optimal production schedules for open pit mines require complex modeling and computational tasks. A realistic model should incorporate mathematical rigors, technical, and economic constraints. In order to find out the sequence of extraction in a predetermined final pit, the models must be able to deal with the limitation of computing resources, time and space. The extensive search domain for production scheduling of a real size mine makes operations research methods virtually impossible to implement. The open pit long-term planning problem is a dynamic uncertain system with a number of underlying stochastic variables, which are not addressed in the current mine planning algorithms.
This paper reviews the development of an agent-based framework for open pit mine planning. The study is a hybrid research work comprising development of an agent-based theoretical framework for large open pit operations, and implementation of the algorithms. The long-term mine planning problem is modelled as a dynamic decision network. The mine production scheduling is formalized as a reinforcement learning problem. The intelligent agent learns the optimal long-term plan by interaction with the block model. The interaction of the reinforcement learning entities: the agent and the block model are modelled by means of stochastic simulation. The simulator captures the open pit layout expansion dynamics over the mine life. The agent employs Q-learning algorithm to maximize the net present value of the mining operation. The developed algorithms of intelligent open pit production simulator, IOPS, are implemented in Java® and MATLAB®.
A case study of an iron ore deposit was carried out to verify and validate IOPS models. The optimized pit limits were designed using Lerchs-Grossman’s algorithm. IOPS annual production schedule was compared to the Milawa production scheduler results, a program widely used in industry. The optimized final pit limits show a total amount of 391 million tonnes of material consisting of 209 million tonnes of ore and 182 million tonnes of waste. The Milawa scheduler yielded an NPV of $430 million over a 21-year mine life at a discount rate of 10% per annum. The IOPS scheduler yielded an NPV of $438 million under the same circumstances and over the same mine life after 3,000 iterations of simulation. The suggested mining starting point of the IOPS schedule was located inside the smallest pit generated by Milawa.
The analyses and comparisons of the results demonstrate that the new planning framework has the potential to substantially improve the expected net present value of mining investments. The intelligent agent framework provides a powerful basis for addressing the real size open pit mine planning problem. Further focused research is required to develop and test the models based on intelligent agents to include more critical mine planning variables such as: optimized cut-off grades, constant annual mill feed, blending parameters, and stockpile constraints into the intelligent mine planning framework. Stochastic simulation as one of the major entities of the developed models has the ability to address the random field and dynamic processes involved in mine planning. The intelligent agent framework has the capability to be used for the optimal integration of mining and mineral processing systems, and development of a framework to quantify uncertainty relevant to mine planning and engineering design.
