Introduction

Video games, since being invented in the 195s0s, have steadily improved at holding the attention of adults and children alike. From the humble beginnings of Pong, video games have changed and evolved, to games with 200 simultaneous players and ray tracing. Video games are as popular as they have ever been, with the stereotype of the “gamer” slowly fading away. One would have trouble finding a person under the age of 30 that had not heard of Call of Duty or Fortnite. The market share of video games was estimated at $217 billion in 2022 and is only projected to grow. The number of people that are able to make a career solely playing video games has risen steadily as well, with streaming sites like Twitch and Kick becoming mainstream.

As the computational power of consumers has increased, the level of product that video game developers produce has scaled accordingly. Video games have grown incredibly complex compared to the predecessors, demanding magnitudes more of storage. However, a small movement has begun to both play and archive older games. As consoles like the Sega Genesis and the Nintendo Entertainment System become older and harder to find, enthusiasts have begun copying the contents of game cartridges onto their computers and sharing them freely across the internet. While the legality and morality of those actions are both questionable, it shows that people today still find some value in playing and preserving older titles.

As these older titles became more freely available, people have begun toying with them. Through use of an emulator, which is a program that reads the memory from the game files and processes it as the original console would, hobbyists have been able to create many awesome things. Communities have sprouted up around older games and done things like writing custom emulators and netcode that allows users to play online against others across the world. Another avenue of interest has come from researchers. In 2016, OpenAI released gym, a “toolkit for reinforcement learning research”. One part of this toolkit included games from the Atari 2600, which was released in 1977.

OpenAI Gym's Atari game environments. (a) Breakout (b) Pong and (c) Video Pinball's RGB frames

OpenAI Gym’s Atari game environments. (a) Breakout (b) Pong and (c) Video Pinball’s RGB frames

One thing about video games that attracted researchers was their deterministic properties. While not all video games are deterministic, researchers found that many of these older games were. Additionally, these games also came with a clear scoring system and objective. Previously, researchers had taken interest in board games like chess or Go. The company DeepMind has spent years developing a program to play Go, culminating in a model named AlphaGo. This model beat the top-ranked human player, Lee Sedol, in 2016. After significant achievements, researchers began developing and working on models to play video games.

Lee Sudol vs AlphaGo in GO

Lee Sudol, playing a game of Go against AlphaGo

At the core of almost any game, is the goal to win. Winning can take many forms, but it is often what drives people to return to a given game. One such example of a popular, and highly competitive game, is Dota 2. Dota 2 is a multiplayer online battle arena game (MOBA) that was released in 2013 and holds a massive playerbase to this day. The game is very complex and is played between two teams of five players. Players that perform far above average can even join professional teams and compete for prizes up to $40 million . OpenAI set their sights on making a model to play as an entire team in Dota 2. In 2019, the OpenAI team met their goal by beating the reigning world championship team in a best of three series. While it was certainly not their first try, with their model having experienced about 45,000 years of play in the game, it marked another milestone for computers playing video games. The researchers also found a new avenue for exploration. They found that their model was able to competently play a match with pro-level human players. This potential of human-AI interaction is something that both players and researchers could greatly benefit from. Many gamers could agree that having a competent, always-available teammate would enhance their experience.