PERFORMANCE OF A MYOPIC LOT SIZE POLICY WITH LEARNING IN SETUPS

Abstract

Abstract We consider the problem of lot sizing when learning results in decreasing setup costs. Finding optimal lot sizes requires information about future setup costs and also the horizon length, which can be difficult to forecast. We analyze an intuitively appealing and well known myopic policy (Part Period Balancing). This policy sets the current lot size such that the current setup cost equals the holding cost for the current lot. It is easy to implement and does not require information on future setup costs. It is shown that the number of setups in the myopic policy is at most one greater than the optimal number of setups. Using this bound, we show that the myopic policy costs no more than 6/(3 + min(l, 1.5R)) times the optimal cost, where R is the ratio of the minimum setup cost to the initial setup cost. Computational experiments show that its average performance is good even for horizons as short as eight times the initial reorder interval. Further, our study shows that the average performance improves with R.