Machine Learning Model for Determination of the Optimal Strategy in an Online Auction

We apply a machine learning model to determine the optimal strategy in an online auction for the rent of computing resources using the best-choice model. The best-choice model allows clients to minimize the expected cost of renting a computing resource based on the spot price distribution function. The spot price dynamics platform is investigated. The most suitable price distributions in an auction are the normal distribution and its mixtures. In this case, the problems of determining the number of components in the mixture and estimating its parameters arise. One of the well-known methods for determining the number of components in a mixture of normal distributions is the BIC criterion. The EM algorithm is a basic tool for estimating the parameters of a mixture of distributions if we know the number of components. However, parameter estimation by this method takes more time when both the sample size and the number of components of the mixture increase. To automate and expedite the process of determining the number of components for a mixture of normal distributions and estimating its parameters, a classification machine learning model based on a convolutional neural network is developed. The results of the model training and validation are presented. The suggested model is compared with other algorithms which do not use neural networks. The results show that the suggested model performs well in determining the most appropriate number of components for a mixture of normal distributions and in reducing the time spent on applying the EM algorithm to estimate its parameters. This model can be used in different arias, for example, in finance or for determination of the optimal strategy in an online auction for the rent of computing resources.

machine learning, auction, best-choice problem, mixture of normal distributions, EM-algorithm