“We directed our attention to the grinding process in Aitik, where we have a well-developed simulation model. We wanted to see if AI was able to do better than our existing control strategy. Mattias did a fantastic job setting up the architecture and getting the various environments to “play ball” with each other. We were then able to test various algorithms and different goal functions,” says Johannes.
To begin, we tested a Q-learning algorithm whose goal was to try to control the mill’s load within a given range. After around 40 attempts, the algorithm had taught itself to do just that. On the other hand, we noted that it solved the task using a method that would not work in the real world. In the next step, we investigated the ability of the algorithm to optimize a gain instead of optimizing a process variable. The goal function for the gain was created as a theoretical model using metal prices, grinding and throughput, for example.
“With this goal function, the AI algorithm succeeded in beating our PID structure to produce a greater gain. So-called wall time was around 80 hours before AI had learned to run the process profitably, in this case equivalent to a plant operating time of more than 300 years.