Team Recommendation Using Order-Based Fuzzy Integral and NSGA-II in StarCraft

As one of the well-known real-time strategy games, StarCraft has become an important benchmark for game artificial intelligence research. Previous works in StarCraft mostly focused on strategic planning and tactical reasoning. One of the key issues in strategic planning, namely unit selection, is to build up an army, so called team in this article, with appropriate units which can gain massive destroy power against enemy's army. It is still a challenge to determine a team that has a good chance of defeating a specified army and there is no any formal algorithm solving it directly at present. Considering that the number of each unit will change during a game, if the player encounters the same enemy for multiple times, we need to select multiple winning troops, which will give the player a number of choices so as to increase the winning chance. In this article, we formulate the team recommendation as a multi-objective optimization problem and propose a novel team recommendation algorithm to solve the problem. We add a normalization of the team size to the order-based fuzzy integral and the normalized order-based fuzzy integral can better estimate the relative combat power of a team. We use genetic algorithm (GA) to learn the fuzzy measure in the fuzzy integral from the StarCraft replay data and adopt Non-Dominated Sorting Genetic Algorithm (NSGA-II) combined with the fuzzy integral for team recommendation. Finally we use a simulator called SparCraft to examine the new algorithm. The experimental results show that our proposed algorithm can recommend winning teams with a high accuracy for ordinary units in StarCraft, and the sizes of recommended teams are mostly not larger than the size of the enemy's team.