An Artificial Intelligence forecast system for strategic positioning in a professional competitive E-Sport

Abstract

The gaming industry is one of the largest contributors in the entertainment industry, boasting an estimated US$152.2 billion of revenue for 2018, with electronic sports (E-sports) being one of the main contributors to this astronomical revenue. One such E-sport is Counter-Strike Global Offensive (CSGO). The purpose of the study is to develop an automatic strategy predictor (ASP) for CounterStrike Global Offensive (CSGO) that will be able to determine a specific team’s strategy accurately. The ASP achieved this by developing an MVS to analyse and capture historical matches and generate a player position stencil (PPS). This PPS is then processed further to obtain the necessary information to train an artificial neural network (ANN). The development of the ASP consists of intercommunicating multifaceted software components. The first part is an image-processing section that captures the in-game mini-map within CSGO and by doing so is able to capture the player positions at a specific point in time for every playable round. This image-processing part consists of two components. The first is the player position stencil producer (PPSP); the second is the stencil processor (SP). The PPSP captures the in-game minimap using the image acquisition techniques and the SP processes the stencils captured by the PPSP and produces the player position information (PPI). An information processing system (IPS) is developed in order to process the PPI. The ISP consists of three components. The first of these was an information concatenator (IC). The IC would concatenate the PPI, captured by the SP, with the additional information obtained from the CSGO Demo Manager. The CSGO Demo Manager is a software package that produces information about every round within a match. The additional information obtained from the CSGO Demo Manager contains information such as the team’s economy, their utility, whether or not the in game device was planted or not and if the round was won or not. This is concatenated the PPI with the additional information, a comprehensive understanding of the round can be generated. Once concatenated by the IC, the information is called the overall round strategy information (ORSI). The second and third software components form part of the information pre-processor (IPp). The purpose of the first IPp software component is to retrieve the ORSI from memory and perform pre-processing on the ORSI. The purpose of the preprocessing of the ORSI is to ensure that the ORSI was in the correct format for the training of an artificial neural network (ANN). The second IPp software component separates and arranges all the information by rounds. The information of each round is separated and stored as the information file for training and is called the overall round strategy information 2 (ORSI2) file. Once each round has an ORSI2, the fourth milestone is met. Each round’s ORSI2 file is used to train an ANN. The ANN of each round is called the automatic strategy predictor (ASP). Once all the ASPs have been trained, five unseen matches are processed by the MVS, the IPS and the IPp in order to obtain five unseen ORSI2 files. These ORSI2 files are used as inputs to determine the accuracy of the ASP when predicting unseen matches. The outputs of each rounds ASP are compared to the actual player positions retrospectively. The results show that the suite of software developed in the MVS was able to capture the in-game mini-map, as well as process the PPS that was captured. The information obtained by processing the PPS was successfully processed into a format that can be used for training an ANN. The training of the ASPs was successful, as the average R values obtained for each round’s ASP, were greater than 0.98. With all the APS’s trained, five unseen matches were processed once more in order to obtain five new ORSI2s. The new ORSI2s were used to determine the accuracy of each round’s ASP when predicting the strategy used by the team within that specific round. The second round’s ASP was able to achieve an accuracy of 67%; this is expected, as there is only one prior round of information for the ASP to make an accurate prediction. The later ASPs all achieved accuracies well over 95% prediction success. The trained ASP delivered instant predictions once an input was given to the ASP and could be utilised by a professional E-sports team to study an opposing team, or by E-sports analysts and discussion panel.