A comparison of distributed and centralised agent based bundling systems

P. Gradwell,J. Padget

DOI: https://doi.org/10.1145/1282100.1282108

2007-08-19

Abstract:The use of trading agents to manage the allocation and bundling of resources across computer networks is well established and literature to date has focused on a variety of auction and distributed market type mechanisms that use economic principles to determine the "best" allocation. An empirical analysis of a number of solver algorithms, principally the Centralised Combinatorial Auction Solver (CASS), has shown that those using bounded search techniques are typically able to solve a majority of cases in linear time, while there remain a number of outlier cases that are computationally problematic. In contrast, distributed mechanisms are intrinsically less than optimal for sellers, but demonstrate signifcantly less variance in computation time. A proper understanding of the different performance properties and suitability of the different techniques is necessary in order to make an informed choice between a distributed market and a centralised auction. Consequently, we have completed an empirical evaluation of CASS, a centralised mechanism, against two distributed mechanisms: (i) Multiple Distributed Auctions (MDAs) and (ii) Quote Driven Markets (QDMs). Uniquely, we carry out simulations of all three mechanisms using a common dataset, generated by the Combinatorial Auction Test Suite (CATS), providing a real basis for comparison. The main results presented are that distributed mechanisms deliver (i) increases in the number of items traded (ii) a greater proportion of bidder requirements being satisfied, but (iii) potentially less optimal bundle solutions and (iv) consistent run times with low overall variance when compared with centralised algorithms.