- The objective of this research project was to identify the limitations associated
with schedule-based Material Requirements Planning (SBMRP) and to
present a knowledge-based expert system (KBES) approach to solve these problems.
In SBMRP, the basic strategy is to use backward or forward scheduling
based on an arbitrary dispatching rule, such as First-In First-Out. One of the
SBMRP weak points is that it does not use such job information as slack times,
due dates, and processing times, information which otherwise is important to
good scheduling decisions. In addition, the backward scheduling method produces
a better schedule than the forward scheduling method in terms of holding
and late costs. Dependent upon job characteristics, this may or may not be
true and should be tested.
This study focused on the means to overcome these two weak points by
the use of a KBES. Heuristic rules were developed through an experiment-based
knowledge acquisition process to generate better schedules, rather than
relying solely upon forward or backward scheduling. Scheduling performance
was measured, based on the minimization of the sums of holding and late costs.
Due to complexities of the problem, heuristic methods were used rather
than analytic methods. In particular, five loading rules were selected, based
upon their close relationship to selected job characteristics, including processing
times and due dates. Combined loading methods (CLMs) were developed to
obtain better performance, derived by combining the two existing SBMRP
scheduling strategies with five loading heuristic rules. This resulted in the generation
of 10 CLMs for further evaluation.
Since this study proposed a new approach, an expert human scheduler
was not available. To overcome this problem, knowledge acqusition through
computer experimentation (KACE) was utilized, based upon an architecture of
five components: job generator, scheduler, evaluator, rule generator (an extended
version of ID3), and the KBES. The first three components were used
to generate a large number of examples required by the rule generator to derive
knowledge. This derived knowledge was incorporated into the KBES.
Experimental results indicated that the KBES outperformed the two
existing SBMRP methods. Based on sensitivity analysis, the KBES exhibited
robust performance with regard to every job parameter except number of parts.
As the number of parts was increased, KBES performance was subject to degradation
since the possibility of interactions or conflicts between parts tended to
increase, resulting in shifting the threshold ratio of total available time to total
processing time. Thus, it is strongly recommended that a new KBES capable of
accommodating 30 parts or more should be developed using the KACE method.