An Uncertainty-Incorporated Approach to Predict the Winner in StarCraft II Using Neural Processes

Mu Lin,Wei-Ping Wang,Jinjun Liu,Yifan Zhu,Xiao-Bo Li,Tao Wang,Yanfeng Wang

doi:10.1109/access.2019.2930581

Mu Lin, Wei-Ping Wang + Show 5 more

Open Access

https://doi.org/10.1109/access.2019.2930581

Copy DOI

Abstract

Real-time strategy (RTS) game is a kind of strategy game, in which the players compete for resources on 2D terrain by establishing the economy, training army, and guiding them into battle in real time. The winner prediction of the RTS games often involves studying a highly uncertain problem in an adversarial environment. In addition, the limit of the number of samples restricts on the application and performance of the prediction models. To obtain better winner prediction accuracy and maintain the prediction uncertainty under an adversarial environment, this paper proposes a neural network-based prediction method incorporated probability inference dealing with a small set of samples. This paper uses a dataset released based on SC2LE, a reinforcement learning environment released jointly by Blizzard Entertainment and DeepMind, and then employed the proposed neural processes model to build a winner prediction model. To verify, this paper implemented different features types' grouping and different game length grouping experiments for demonstrating better adaptability to such problems. Furthermore, this paper also implemented the SVM model experiments and compared the proposed method with the SVM model. Finally, when making predictions on a 1000 size testing data, the results show that the proposed prediction model achieves an accuracy of 0.811 at 200 and 0.821 at 1000 sizes of training sets, which is better than the SVM model with small training datasets.

Highlights

Real-Time Strategy (RTS) is a sub-genre of strategy games, in which players compete on a 2-D terrain, by building a base, gathering resources, training units, and guiding them into battle in Real-Time [14]
It can be seen that the prediction accuracy of the SVM model for short and middle length grouping is significantly improved compared with that of ungrouping, while the prediction accuracy of Neural Processes (NPs) model is not significantly improved after grouping
The NPs neural networks model was developed to estimate the winner of the game through adversarial features and non-adversarial features

Summary

Introduction

Real-Time Strategy (RTS) is a sub-genre of strategy games, in which players compete on a 2-D terrain, by building a base, gathering resources, training units, and guiding them into battle in Real-Time [14]. The RTS game’s prediction of outcomes (Win or Defeat) is an interesting area for Artificial Intelligence (AI) research. It is an effective environment, to conduct experiments for complex adversarial systems in RTS games [19], and the prediction of winners is done via depiction as a typical, large-dimensional, non-linear, probabilistic inference problem. Players are expected to carefully adjust their strategies, based on a large number of observable dynamic variables, such as resources, supplies, units, and other factors, that will be referred to as features. A lot of features involved are completely random or unobservable, during the preceding game-play, which may have a major impact on a game’s outcomes.

Objectives

Results

Conclusion