GBDT, LR & Deep Learning for Turn-based Strategy Game AI

EasyChair Preprint no. 1053

Abstract

This paper proposes a AI fighting strategies generation approach implemented in the turn-based fighting game StoneAge 2 (SA2) [1]. Our research aim is to develop such AI for choosing the logical skills and targets to the player. The approach trained the logistical regression (LR) model and deep neural networks (DNN) model, individually. And combined both output at inference process. Meanwhile, to transform the features into a higher dimension binary vector without any manual intervention or any prior knowledge, we put all category features into Gradient Boosted Decision Tree (GBDT) before LR component. The main advantage of this procedure is: the approach combines the benefits of LR models (memorization of feature interactions) and DL (generation the unseen feature combination through low-dimensional dense feature) for the AI fighting system. In our experiment, we evaluated our model with some other AI strategies (Reinforcement Learning (RL), GBDT, LR, DNN) to against a robot script. The results shown that the players, participating in the experiment, are capable of using reasonable strategic skills on the different targets. As a consequence, the win rate (versus with the robot script) of our system is higher than the others. Finally, we productionized and evaluated the system on SA 2, a commercial mobile turn-based game.

Keyphrases: deep learning, Fighting Game, GBDT, logistic regression, machine learning, Recommender System, Stone Age, turn based battle recommendation pipeline, Turn based Game AI

@Booklet{EasyChair:1053,
  author = {Like Zhang and Hui Pan and Qi Fan and Changqing Ai and Yanqing Jing},
  title = {GBDT, LR & Deep Learning for Turn-based Strategy Game AI},
  howpublished = {EasyChair Preprint no. 1053},

  year = {EasyChair, 2019}}