Production constrained build-order optimization problems challenge artificial intelligence research in computer game applications due to an uncertain set of constraints. Traditional approaches provide subjective values in the constraint formulation therefore resulting in unexpected performance of the optimal build-order in a game. In this article, we propose a data-driven approach to solve a build-order optimization problem in StarCraft. We formulate the constraint by learning the parameter values from game replay data, which complements more precise problem formulation. To solve the optimization, we use the improved genetic algorithm by learning initial solutions from the data. We show the performance of the data-driven methods in a StarCraft simulation platform.
|Name||Chinese Control Conference (CCC)|
|Conference||2019 Chinese Control Conference |
|Period||27/07/19 → 30/07/19|