Reinforcement learning for grinding – New publication!

About a year ago, me and my colleague Mattias Hallén started investigating the use of reinforcement learning in controlling grinding circuits in mineral processing.

The work has been done in collaboration with the mining company Boliden, and it has currently been showcased at various avenues:

  • IEEE Conference on Emerging Technologies and Factory Automation 2019 in Zaragosa, Spain
  • The 2019 Mining User Conference in Santiago, Chile
  • It was featured both in International Mining and the Mining Magazine.

The very short summary is that we have used reinforcement learning (Proximal Policy Optimization) to control a grinding circuit in simulation. This is done by using OpenAI, and interfacing with a Dymola-simulation by compiling the simulation model into a Functional Mock-up Unit. The resulting controller was competitive with existing PID-based control.

The grinding circuit