OpenAI starts machine learning competition with classic Sonic games

OpenAI, which deals with the development of artificial intelligence, has announced a so-called Retro Contest. Entrants must submit an AI agent who will then play unknown custom Sonic levels.

According to OpenAI, the competition is aimed at evaluating a reinforcement learning algorithm, a form of machine learning. This is done by looking at how it can generalize based on previous ‘experiences’. Instead of using a familiar environment for training as well as for the competition itself, participants will have access to a training set of levels from various Sonic retro games. During the competition, their agent has to pass through levels unknown to the developers within a time frame of eighteen hours per level. OpenAI has also published the benchmark to be used.

The competition lasts two months and runs until June 5. The winner is the one with the highest score in the ranking. Participants can submit their agent in the form of a Docker container. OpenAI itself has released a number of basic algorithms that participants can use as a starting point. In addition, the organization has made available the Gym Retro Beta environment, which can be compared to a kind of emulator for the Atari 2600 and Sega Genesis that can be used to use games as training environments. OpenAI has used this type of Gyms for training with games before.

In the announcement, OpenAI also mentions that agents sometimes find exploits in games, depending on how the rewards function is set up. For example, one of his agents discovered that it was possible to glitch through a wall and achieve a higher score.