A Self Organizing Bin Packing Heuristic

Abstract

This paper reports on experiments with a new on-line heuristic for one-dimensional bin packing whose average-case behavior is surprisingly robust. We restrict attention to the class of “discrete” distributions, i.e., ones in which the set of possible item sizes is finite (as is commonly the case in practical applications), and in which all sizes and probabilities are rational. It is known from [7] that for any such distribution the optimal expected waste grows either as θ(n), θ($\sqrt n$), or O(1), Our new Sum of Squares algorithm (SS) appears to have roughly the same expected behavior in all three cases. This claim is experimentally evaluated using a newly-discovered, linear-programming-based algorithm that determines the optimal expected waste rate for any given discrete distribution in pseudopolynomial time (the best one can hope for given that the basic problem is NP-hard). Although SS appears to be essentially optimal when the expected optimal waste rate is sublinear, it is less impressive when the expected optimal waste rate is linear. The expected ratio of the number of bins used by SS to the optimal number appears to go to 1 asymptotically in the first case, whereas there are distributions for which it can be as high as 1.5 in the second. However, by modifying the algorithm slightly, using a single parameter that is tunable to the distribution in question (either by advanced knowledge or by on-line learning), we appear to be able to make the ratio go to 1 in all cases.

Supported in part by a Fullbright Fellowship, by DIMACS, and by AT&T Labs

Supported in part by AT&T Labs