The p-strategy: An adaptive agent bidding strategy based on stochastic modeling for continuous double auctions

We expect an e-commerce infrastructure to be populated by software agents who buy and sell goods on behalf of their human owners. As a step towards making this vision come true, we have developed an agent bidding strategy (called the p-strategy) for a class of continuous double auctions (CDAs).; The p-strategy is based on stochastic modeling of the auction process. A p-strategy agent uses probabilistic assessment about offer price distributions, offer arrival rates, etc. to figure out the best price to offer at the auction. Using its stochastic model of the auction, the p-strategy agent trades off the probability of success against the payoff of success, which is manifested in its decision of raising (or dropping) and narrowing (or spreading) its offer prices.; We have evaluated the performance of the p-strategy through experiments. We have divided the experimental space into three dimensions, and systematically compared the p-strategy seller to sellers with different bidding strategies in various environments. We have found that the p-strategy outperforms other agent strategies in the CDA in a majority of experiments. In particular, the p-strategy seller performs well when many buy offers are available, as it can extract more profit per match at the expense of buyers. However, the performance of the p-strategy seller degrades with high competition among sellers or with multiple competing p-sellers.; In addition to stochastic modeling, we have developed an adaptation algorithm. Although the p-seller performs very well in most cases, there are some cases where the sophisticated modeling of the auction does not pay off. In addition, stochastic modeling requires non-trivial computation, and thus deliberation overhead diminishes the advantage of using the stochastic model. The adaptation algorithm allows the p-seller to adaptively figure out when to use stochastic modeling or not at run time. Adding adaptivity to the p-strategy is an important and practical step for using the p-strategy. The experimental results indicate that the adaptive p-strategy outperforms the plain p-strategy when the p-strategy performs poorly, while it performs very similarly to the p-strategy when the p-strategy dominates other simple strategies.