League of Legends is a multiplayer game, which needs tacit teamwork. How to make the right decision is very important in the ever-changing battle. Recently, Philip Osborne, a data analyst, proposed a method to improve the decision-making level of League of Legends team by using artificial intelligence technology, and made it open. This method not only refers to the statistical results of a large number of real games, but also considers the preferences of current players.
The project consists of three parts, aiming at modeling the battle of MOBA game League of Legends as markov decision processes, and then applying reinforcement learning to find the best decision, which also takes into account the player's preferences and goes beyond the simple "scoreboard" statistics.
The author uploaded various parts of the model in Kaggle, so that everyone can better understand the data processing flow and model structure:
Part I:/Osborne/lol-ai-model-part-1-initial-EDA-and-first-MDP Part II: /Osborne/lol-ai-model-Part-2- redesign-MDP with gold-diff Part III:
At present, this project is still in progress, and we hope to show what complex machine learning methods can do in the game. The score of this game is not just a simple "scoreboard" statistical result, as shown in the following figure:
Motivation and goal
League of Legends is a team competition video game. There are two teams (five people in each team) in each game to recruit and kill. Gaining an advantage will make players stronger than their opponents (better equipped and faster upgraded). If one side's advantage continues to increase, the chances of winning will also increase. Therefore, the subsequent style of play and the direction of the game depend on the previous style of play and the war situation, and the last party will destroy the opponent's base and win the game.
Modeling based on precedent like this is not new; For many years, researchers have been thinking about how to apply this method to basketball and other sports. This article explains how to map feedback in more detail.
How to collect feedback determines how successful our model can be. In my opinion, the ultimate goal of doing this is to provide the best real-time advice for the player's next decision. In this way, players can choose from several best decisions (ranking according to the winning situation) calculated according to the game data. You can track players' choices in multiple games to further understand and understand players' preferences. This also means that we can not only track the result of the decision, but also predict the player's intention (for example, the player tried to dismantle the tower but was killed), and even provide information for more advanced analysis.
Of course, such an idea may cause differences among team members, and it may also make the game less exciting. But I think this idea may be beneficial to low-level or regular-level players, because it is difficult for them to clearly convey their game decisions. This may also help to identify the "Cancer" players, because the team hopes to unify their opinions through the voting system, and then it can be seen whether the "Cancer" players have not acted according to the team plan and neglected their teammates.
Example of Model Recommendation Voting System in Real-time Game Environment