Using Reinforcement Learning for Generating Polynomial Models to Explain Complex Data

Abstract Basic oxygen steel making is a complex chemical and physical industrial process that reduces mix of pig iron recycled scrap into low-carbon steel. Good understanding the ability to predict how it will evolve requires long operator experience, but this can be augmented with target prediction systems. Such systems may use machine learning learn model based on history, have an advantage in they make vastly more parameters than operators comprehend. While has become less challenge build such using algorithms, actual production implementations are rare. The hidden reasoning lack transparency prevents trust, even for models show high accuracy predictions. To express behaviour thereby increasing we develop reinforcement (RL) agent approach, which task generate short polynomials explain from what learnt data. RL rewarded well generates smaller subset parameters. Agent training done REINFORCE algorithm, enables sampling multiple concurrently plausible polynomials. Having polynomials, developers evaluate several alternative explanations, as observed historic presented approach gives both trained generative set process. performance good or better interpretable models. Further, relative simplicity resulting allows generalisation fit new instances best our evaluation achieves $$R^2$$ R2 score test comparison other evaluated.