Information Technology Reference
In-Depth Information
2.50E+07
2.25E+07
2.00E+07
1.75E+07
1.50E+07
1.25E+07
1.00E+07
7.50E+06
5.00E+06
2.50E+06
0.00E+00
24780
31860
38940
46020
53100
60180
67260
74340
81420
Total Day 0
Fig. 1. Profile data for storage costs of 100% annually. The plot shows the -bound for explored
profiles against the aggregate level of day-0 purchasing (per supplier/component) for all strategies
in the profile. The dark box represents the region for which there are known profiles with -bounds
less than $2.5M.
setting (playing 5-10 games for each profile). We then performed 12-32 iterations of
the best-first search procedure for each setting of storage cost. We ran a total of 2670
games over 6 months, exploring approximately 10% of the total profile space for these
discrete parameter and strategy settings.
Figure 1 shows a plot of the data for annual storage costs of 100% (the mean storage
cost setting from the 2004 tournament). Each point represents the -bound for a sam-
pled profile, plotted against the aggregate day-0 procurement in the profile. To calibrate,
a total procurement of 35400 (total multiplier 3.0) corresponds to an expected commit-
ment of approximately 1/3 of the total supplier capacity for the entire game. Note that
many different profiles have the same aggregate procurement. The dark box shows the
region with the most stable (lowest- ) profiles. This region yields a predicted range for
the total day-0 procurement induced by this storage cost setting.
Figure 2 shows results for a range of settings of the storage cost parameter. The
SearchMin and SearchMax lines correspond to the endpoints of the region defined like
the gray region in Figure 1. The other three lines indicate approximate equilibria found
by the three learning methods, trained on the initial 10 randomly-generated profiles for
each storage cost setting. It is encouraging that the results obtained using very differ-
ent methods (learning and directed search) have the same qualitative structure. This
experimental evidence supports the initial intuition that day-0 procurement should de-
crease with higher storage costs; all of the methods show this relationship. There is
Search WWH ::




Custom Search