Knowledge acquisition for discrete event
systems using machine learning
Knowledge acquisition for an understanding of discrete event simulation
systems is a difficult task. Machine Learning has been investigated to help
in the knowledge acquisition process. Our approach involves consultation
with a domain expert, and the use of discrete event simulation models and
machine learning as tools for the intelligent analysis of simulated systems.
Current methods for the analysis and interpretation of such systems are
restricted to statistical techniques that say much about the reliability of
an output, but little about the output inter connectivity.
The objective of our
work is to improve the ability to interpret the model to the level of
explanation that might loosely be described as ``How the simulated system
works''. The new interpretation techniques are based on knowledge
acquisition concerning the system using machine learning tools.
``How the simulated system works'' is important for a decision maker's
understanding of
the system in terms of relationships between the various parameters, the
utilisation of resources and the location of (potential) bottlenecks. This
sort of understanding is important when managing queuing problems in
production systems, or when planning a new system.
It should be emphasised that the goal of the presented work is not a
prediction nor an optimisation. The main goal here is to find interpretations
of the simulation output and discovering regularities, thus helping
the user to develop
an intuitive understanding of the domain.
Discrete event simulation produces example situations that can be used
as input data for machine learning tools.
In the presented research, three simple and commonly used discrete
event simulators were interpreted using different machine learning tools.
The attribute based learning systems RETIS and ASSISTANT
were chosen as appropriate for learning in noisy, real-world domains.
The interpretation obtained by these systems was intuitive but obviously
expressed in a complicated way. To enable a more powerful knowledge
representation, the Inductive Logic Programming (ILP) system
MARKUS was used. In addition to compact knowledge
representation, MARKUS also highlighted some attribute combinations
that could be useful in the attribute based learning.