Pedro Haschelevici
Minerva University - 2023
Gaming is a fairly new part of the entertainment industry, with the release of Pong (considered one of the first video-games) having its 50th anniversary this year. Even at this early age video games are already an indispensable part of our lives, with over 2.5 billion people playing them worldwide, and an expected global market of 257 billion dollars by 2025 (How Much Is the Gaming Industry Worth in 2022? [+25 Powerful Stats], 2022).
The gaming industry has rapidly become one of the most lucrative parts of the entertainment industry, growing at an exponential rate both in terms of technological advances and overall revenue. The global gaming market was worth more than $100 billion in 2018, and this number is only expected to increase as the industry continues to expand. In addition to the traditional console and PC gaming markets, gaming has also found an important niche in mobile gaming, as well as esports and streaming platforms, which have become immensely popular over the past few years. As gaming continues to evolve and diversify, it is clear that this industry is here to stay, and will only continue to grow in the years to come.
One genre of video-games that particularly interests me is Roguelike. They are games in which you have to restart your gameplay from the beginning every time you die, and a crucial aspect of it includes randomly generating the levels so each playthrough feels unique. (Master Class, 2021) This randomness, however, is very tricky to get right. Too random and the runs feel disconnected and pointless, too little random, and they all feel the same. This problem could be avoided if the video-game “knew” who is playing the game and adjusted its levels based on how the player is playing. However, we can’t, unfortunately, have an unlimited number of pre-generated levels for every type of player in every playstyle, there is just not enough resources for that to be made. What we can do, however, is try to adjust the randomness of the levels.
This could be done by using a Bayesian process, which would allow the game to “learn” as the player plays and adjust the randomness of the levels accordingly. In practice, it would identify patterns in the player’s playstyle and then adjust the randomness of the levels to fit the player’s style. This would allow for the game to be unique for each individual player, while still offering the same amount of challenge. This could be used to create a much more fluid experience that would offer a lot of replayability.
Moreover, this same process could be used to create dynamic difficulty levels, making the game easier or harder depending on the player’s progress. This would allow for a much more personalized experience, as the game would be different every time the player plays. Overall, Roguelike games are incredibly interesting and fun, and the use of a more sophisticated random level generation could make them even better.
Random level generation is made based on parameters and distributions. There is a set of attributes present in a level, like the number of rooms, size of the rooms, number of enemies, type of enemies, damage of the enemies, location of hidden chests, loot drops, and many others. These attributes are generated using distributions, maybe the damage of a skeleton is defined by a normal distribution around a value, or the number of rooms can be simply an uniform distribution between a minimum and maximum values. By adjusting these distributions we can drastically change how the levels are generated, and an interesting question comes with this: Why do these distributions have to be the same for every player? Well, short answer, it doesn't have to be. Personalizing these distributions might be the key for also personalizing the level generation, giving the player unique, but still consistent and tailored levels every run. This is what I’ll try to achieve with this project.
I want to make a roguelike game that delivers personalized random levels. In order to achieve this, I plan to use a more data driven approach. The game will collect data on how you play the game and build a kind of a player profile. This profile will control the distributions using a Bayesian approach, where we have base distributions (our prior), and as more data about the players play style comes in it generates likelihoods, combining them to get posterior distributions on how the next level will be generated. This process repeats itself throughout the various playthroughs making the level generation more accurate to the player the more they play the game.
While this sounds like a valid approach in theory, in practice, it feels more like overdoing the problem. So, instead of building a complex Bayesian update system for selecting the level, it is much more reasonable for a videogame to have a simple weighted level selection with a big pool of levels to choose from. With this, I can achieve a nearly as good of an experience, without the cost (both computational and time) of building a more complex model.
I decided to divide the process for this in 3 sections: Persona Identification using aspects of Human Computer Interaction, identification of behavior patterns of these personas, and to wrap both of these I’ll be using the DDBT framework. This seemed like the a good approach for me given that I’m taking a game development tutorial this semester, which will give me a good opportunity to gain feedback and insights on different stages of the development of the game through participatory methods, in which the other participants of the tutorial will be playing the game and noting down which aspects they thought worked well and which didn’t.