Many events can be predicted before they happen, for example, it is quite obvious that the character of a person playing the popular MOBA game Dota 2 will soon die if a stronger enemy hero approaches him from a zone out of his sight. But what is obvious to a person is not always easy for a computer, and a person is not always able to follow everything that happens on the game map. IN titled "Time to Die: Predicting the Death of a Dota 2 Character Using Deep Learning," researchers at the University of York described how they were able to train an AI to accurately predict a player's impending death 5 seconds before it actually happens.
In fact, predicting that a character will be killed in 5 seconds is a little more difficult than it seems at first glance. An average match consists of 80 individual fragments, during each of which the character can perform dozens of 000 possible (according to the researchers' calculations) actions. On average, players on the map perform 170 moves per match fragment, making over 000 game changes.
The authors of the study note that the low health of a character is not always closely related to his quick kill, as some heroes have healing abilities, and there are special items for healing or teleportation. Considering all these factors, the team used to train the neural network Dota 2 match records provided by Valve, which contained 5000 professional and 5000 semi-professional games played up to December 5 last year. Before the actual training, the recordings were pre-processed by converting the matches into timelines for each player, divided into 0,133 second slots of game time, where each point on the timeline contained a complete set of data about the character and his environment.
From all the in-game information, the researchers identified 287 parameters, such as health, mana, strength, agility and intelligence of the character, the activated items he has, ready to use abilities, the position of the hero on the map, the distance to the nearest enemy and the defense tower of allies, and See also general review history (when and where was the last time the player saw an opponent). These parameters, as the researchers point out, play a key role in whether the character will die or survive in the short term, while the position on the map and the history of the review played the most significant role.
“The behavior of the players depends on information about the recent past,” write the co-authors of the work. “For example, if the enemy just disappeared from view, the player still knows that he is somewhere in the area. On the other hand, if an enemy disappeared a few minutes ago, it could be anywhere from the player's point of view. This was the reason why we added a feature that analyzes the review history.”
To train the neural network, the scientists used 2870 inputs (287 parameters per 10 players) and 57,6 million data points, reserving 10% of the data for verification and another 10% for testing. Through experimentation, the team found that it achieved an average accuracy of 0,5447 in situations where the AI was asked to predict which of the ten players on either team's hero would die within the next five seconds. In addition, the researchers point out that the model could predict deaths over a longer period of time by studying all the factors and situations that can lead to them.
The scientists note that their approach has certain limitations, namely that the system requires so much in-game information (including about enemy champions invisible to the champion in question) to make its prediction, and that it may not be entirely compatible with new versions. games. However, they believe that the model they have developed, which is available in , can be useful for commentators and players when they follow the match.
“Esports games are very complex, due to the high speed of the gameplay, the balance of the game can literally change within a few seconds, while various events can occur in many areas of the game map at the same time. They can happen so quickly that commentators or spectators can easily miss an important moment in the game and then only observe its consequences, ”the researchers write. “At the same time, in Dota 2, killing an enemy hero is a key event that interests both commentators and viewers.”
Source: 3dnews.ru