https://www.kaggle.com/c/pubg-finish-placement-prediction Our team has chosen the Kaggle challenge of predicting player placements in the Player Unknown Battle Grounds (PUBG) video game. The project link is given above. By using the algorithms and techniques learned in this class, we can use various methods to determine the relevance of different aspects of the player statistics, and decide on how different features affect the scoring. A possible challenge would be that the players performance is not strictly linked to either pure skill or their in-game statistics, and instead has an immeasurable and more random quality. We hope to show that this is not the case and that the players’ observable statistics have a large bearing on performance.

Machine learning techniques are appropriate for this data because we must predict the final placement of players using their statistics gathered in the game. We must see how different features such as the number of eliminations, the amount of boosts used, etc impact the possibility of a player winning. Noting patterns manually given a small enough set of data is personally intensive, and the accuracy becomes limited to the size of training set a person is able to analyze. However, a machine learning algorithm begins in the training effort how we do, by knowing nothing about these patterns, then gaining knowledge in a measurable way, and can eventually come to classify the data to a notable accuracy.

The dataset is available on the main Kaggle project page, under the “Data” tab. We are provided with two sets of data; the training set, and the test set, which contain a large number of anonymous PUBG game stats from around 65000 games, formatted so that each row contains one player’s post-game stats. The data comes from matches of all types: solos, duos, squads, and custom; there is no guarantee of there being 100 players per match, nor at most 4 players per group. So while a player may choose different strategies to play in solo games compared to games with 4-person teams, we must account for all data equally in order to determine common elements of success for all of these game modes. The data itself is provided by the PUBG Developer API.

Our goal is to implement a Multi-layered Neural Network for this dataset. This type of neural net requires minimal pre-processing, making it optimal for the large amount of data and is able to accommodate for non-linear trends in the data. We will start off with a simple, single layered Neural Network baseline and train our data using that. Once we are able to do that, we intend to build our MNN from the baseline. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.

This project intrigued us mainly because we are involved in this gaming community. It seemed an interesting idea to apply the machine learning techniques we have discovered in this class, and to explore statistics for predictions that we otherwise would not have thought about. The idea to predict how a player will place is a very good idea, and can have a great impact on competitive matches where professional gamers can train using the data provided by this exploration. They can, for example, see where their skills/play-style is lacking and improve on that, since there will be evidence to suggest that improving on a certain feature heavily increases the likelihood of winning. This can be also used internally by the PUBG Development team if they were ever to introduce an actual ranking system where the player base is grouped into tiers based on their performance and their predicted placements. In addition, if factors such as boosters and other health regenerating items were the soul determiner of success in a match, it may be worthwhile for the developers to reduce the impact of such items so as to create a more even playing field among the players. We also realised from the start that this operation is very time expensive, and there are a lot of steps needed to be done before we can even begin implementing our algorithms.