Output analysis for stochastic simulation has traditionally focused on obtaining statistical summaries of time-averaged and replication-averaged performance measures. Although providing a useful overview of expected long-run results, this focus ignores the finer behaviour and dynamic interactions that characterise a stochastic system, motivating an opening for simulation analytics. Data analysis efforts directed towards the detailed event logs of simulation sample paths can extend the analytical toolkit of simulation beyond static summaries of long-run behaviour. This thesis contributes novel methodologies to the field
of simulation analytics. Through a careful mining of sample path data and application of appropriate machine learning techniques, we unlock new opportunities for understanding and improving the performance of stochastic systems.

Our first area of focus is on the real-time prediction of dynamic performance measures, and we demonstrate a k-nearest neighbours model on the multivariate state of a simulation. In conjunction with this, metric learning is employed to refine a system-specific distance measure that operates between simulation states. The involvement of metric learning is found not only to enhance prediction accuracy, but also to offer insight into the driving
factors behind a system’s stochastic performance. Our main contribution within this approach is the adaptation of a metric learning formulation to accommodate the type of data that is typical of simulation sample paths.

Secondly, we explore the continuous-time trajectories of simulation variables. Shapelets are found to identify the patterns that characterise and distinguish the trajectories of competing systems. Tailoring to the structure of discrete-event sample paths, we probe a deeper understanding and comparison of the dynamic behaviours of stochastic simulation.