Assessing Usage Patterns to Improve Data Allocation via Auctions

The data allocation problem in incomplete information environments consisting of self-motivated servers responding to users’ queries is considered. Periodically, the servers use auctions for allocation of new data items, and for reallocation of old data items. The utility of a server from storing a data item strongly depends on the usage of the item. However, each server has information only about the past usage of the data stored locally, but does not have information about the usage of data stored elsewhere. In this paper we propose that in order to improve the behaviour of the servers in the auctions, each server learns the expected usage of data items from information about past usage of its own data items. We implemented this type of learning process using neural networks. Simulations showed that our learning methods improve the results of the bidding mechanism, and they are better than the results obtained when learning via k-nearest neighbors algorithms.