Maskinavdelningens lunchseminarium: Enhancing the realism of event outcomes in discrete event simulations through machine learning

02 oktober 2023 12:00 – 13:00 Hörsal T, Teknikhuset, Campus Örebro

Lunchseminarium med Sean Reed.

Discrete event simulation (DES) is an operational research method for modelling stochastic, dynamic systems where the system evolves through a sequence of events and each event occurs at a precise instant in (simulated) time and results in a change to the system state. It is one of the key methods of Industry 4.0, where it is used to form “digital twin” models capable of predicting and optimising the future trajectories of the real industrial systems.

In this presentation, it is shown how supervised machine learning methods can be integrated with DES to enable digital twins that automatically learn from data on the real system, collected from sensors and other sources, to improve predictive accuracy and relevance over time, including under circumstances that were not seen previously or unforeseen at model design time.

Supervised learning is a subset of artificial intelligence where a computer learns the generalized relationship between inputs and outputs from a set of labelled training examples. Event outcomes in a DES are typically random and can be multivariate, for example the event of a part processed by a station in a manufacturing process may be associated with a random processing delay and destination downstream station. Typically, the modeler specifies probability distributions for each event outcome at model design time, which are then sampled from during simulation.

In the pioneering approach presented here, multivariate event outcome distributions are instead predicted during simulation, based on the current system state, prior to sampling of the multivariate outcomes. A general procedure for integrating supervised learning with DES in this way is described. Its application to the modelling of industrial systems using two very different supervised machine learning techniques, namely mixture density networks and distributional random forests, are also given along with a summary of the advantages and disadvantages of these alternative approaches.