Reinforcement learning control of a SAG mill grinding circuit: first impressions and implications for process control, F. Reyes, M. Hilden, M. Yahyaei, and G. Forbes

Since AlphaGo defeated the best human player in the game of Go in 2016 - something thought to be an impossible achievement - machine learning, deep learning and reinforcement learning have caught the attention of scientists, the media and the general public alike. This recent development in machine learning (let us remember that neural networks have been available for decades now) opens new possibilities in the areas of modelling and process control. However, little has been reported in the minerals processing industry about these new possibilities and opportunities. This paper presents and introduces the topic of reinforcement learning and the different platforms that allow its implementation. The paper shows the consequences of using reinforcement learning for the specific area of process control and optimisation. As an example, the authors present the use of reinforcement learning for the control of a SAG mill closed with a pebble crusher using industrial data from an operation. A comparison with advanced model-based process control is provided, showing the benefits and shortcomings of reinforcement learning as a new control and optimisation technique. Keywords: Reinforcement learning, deep learning, process control, process optimisation