Abstract
Decision trees are some of the most popular and intuitive classification techniques. Based on the recursive division of the data, the goal is to ultimately identify regions in the space in which most instances belong to the same class. This paper proposes a game-theoretic decision tree using a two-player game to determine the splitting hyperplane at the node level based on the Nash equilibrium concept. The entropy on each sub-node is used as a payoff function that has to be minimized. The game’s equilibrium can be computed by minimizing an objective function constructed based on Nash equilibria properties. A new selection mechanism is proposed for the Covariance Matrix Adaptation – Evolution Strategy (CMA-ES) in order to approximate equilibria at each node level. Numerical experiments illustrate the behavior of the approach compared with other decision trees based methods.
Citare
@Inproceedings{Lung2022AGT,
author = {R. Lung and M. Suciu},
booktitle = {Eä},
title = {A Game Theoretic Decision Tree for Binary Classification},
year = {2022}
}