Recently, in South Africa, researchers let two AIs have their way in the popular “sandbox” video game Minecraft. The objective? Observe their evolution in this environment and better understand the potential shortcomings of AI in general when they are not guided by humans.
Special AI model for Minecraft
As a reminder, Minecraft is a video game released in 2011. This open world integrates a crafting system focused on the exploitation and transformation of natural resources. is the objective Build structures using small cubes. Minecraft surpassed the 200 million copies mark in mid-2020 for a community of over 125 million monthly active players. Thus it is the best selling game in history.
For Steven James, a researcher at the University of the Witwatersrand (South Africa), Minecraft could be one. A wonderful playground for young AI. The intelligence involved here is truly multitasking, and learning is key to their formation and evolution. So this is a question of machine learning, a process that allows AI to learn from its mistakes and gain influence when solving tasks. without initially programming for.
The AI model from South African researchers is none other than MinePlanner, Made specifically for testing in the Minecraft game. The results of the study were then pre-published on the arXiv platform on December 20, 2023.
Very low performance
The researchers created two AIs using MinePlanner. Each had to ascertain the total 45 constructions divided into three difficulty levels : Easy, Intermediate and Difficult. For the AIs, the most complex thing was dealing with the massive amount of information contained in the game. They then had to discard data deemed unusable for the realization of their constructs and thus progress on their own.
The results were not really conclusive. Indeed, the first AI managed to complete fourteen simple constructions and three intermediate level constructions. The second AI managed to complete only five simple and only one intermediate construction. According to scientists, the second AI especially suffers and His memory falters on most tasks Before you can make them.
The authors explain that there are still technical gaps to overcome in solving some of the problems. However, such work can help Develop new approaches And perhaps in more complex areas. Thus, the future will tell us whether it will be possible to better understand AI learning using the MinePlanner tool.
