Experimental investigation of shared storage assignment policies in automated storage/retrieval systems

IIE Transactions, August, 1999 by Sadan Kulturel, Nur E. Ozdemirel, Canan Sepil, Zafer Bozkurt

The last two rows of the table are different in the manner of rack partitioning. In the third row, racks are partitioned directly by prespecified proportions, either 20/40/40 or 10/45/45 depending on the column. In the second row, however, rack partitioning is dependent on the order quantities. In cells (2,1) and (2,2), rack space allocated to class A is as much as [Q.sub.1] [Q.sub.2]. In cells (2,3) and (2,4) this space is ([Q.sub.1] [Q.sub.2])/2. We try these alternatives to facilitate a fair comparison of the two policies by making the number of units assigned to classes proportional to rack partitioning. For example, cell (2,1), satisfies this property for TCL, and cell (3,2) for DCL.

According to the analysis of variance results, the differences in travel time among the cells of Table 4 are statistically significant at [alpha] = 0.01. The trends, which we discuss below, are the same for both levels of the ABC factor considered in the experiment.

If we examine the travel times given in the table, we see that TCL still outperforms DCL in all cases. On the average, TCL yields 3.61 and 3.16% shorter travel times for ABC factor levels of 20/80 and 20/60. On the other hand, when we run our simulation model with the DOS classification technique described in Section 4, DCL outperforms TCL. This, however, is not a fair comparison for TCL, because in this approach we assign exactly 20% of the items to class A in TCL whereas in DCL more than 20% of unit loads are in this class. Squeezing class A unit loads into 10% of the rack space closest to the I/O point results in shorter travel times in both storage assignment policies. We can say that both policies are highly sensitive to this factor.

These results are in conformity with Goetschalckx and Ratliff's [1] findings. Their experimental results indicate that a strictly DOS-based policy where unit loads are not classified into ABC classes yields the largest savings in AGV travel time compared to DED. However, when a classification scheme similar to ours is employed, two class TCL outperforms two class DCL. They also point out that DCL is very sensitive to the size of the zones and DOS boundaries used in classification of unit loads.

As far as the effects of the (Q,r) policy are concerned, we can say that using Q instead of individual [Q.sub.i]'s is not particularly disadvantageous. However, our Q is not an arbitrarily selected order quantity; it is the average of [Q.sub.i]'s. As a matter of fact, when we have tried larger arbitrary Q values such as 40, we have observed a substantial increase in the AS/R machine travel time. Hence, using the (Q,r) inventory policy is advantageous compared to using arbitrary order quantities even when we ignore the differences among individual order quantities.

7. Conclusion

In this research, we have compared three class turnover-and DOS-based storage assignment policies in an AS/RS environment by means of computer simulation. The AS/RS operates based on a continuous review (Q,r) inventory model used for replenishment of items.