New Scientist magazine recently unveiled on its site a study conducted on Minecraft by a team of researchers from South Africa’s University of the Witwatersrand. The study involved immersing the AI in a game world without any input to test its ability to adapt to its environment.
Video games now occupy a special place in our lives and in our daily lives. So what about Minecraft, the best-selling game of all time with over 300 million copies sold? The answer is simple. According to a magazine article The new scientist, the title is so intense that a team of researchers is using it to test the adaptability of its artificial intelligence. Indeed, there is a team led by him Steven James no’University of the Witwatersrand in South Africa. This study could be key to the development of AI.
MinePlanner, an AI training program for Minecraft
A program developed by researchers is called MinePlanner. So, what does this study entail? The goal is to test the AI’s ability to complete tasks on its own and in the same way that a human can. MinePlanner Measures AI’s ability to sort information. In other words, AI must be able to ignore extraneous details to solve a complex, multi-step problem. According to Steven James :
Many AI training programs “cheat” by giving a model all the data it needs to learn how to do a job and nothing left out. That’s an approach that pays off if you want to build software to accomplish a specific task, like predicting the weather or folding a protein, but not if you’re trying to build artificial general intelligence, or AGI.
James Hope that MinePlanner Can help you tackle complex puzzles. For now, the purpose of the test is to test the AI’s ability to observe the game environment to determine what is useful and what is not. To do this, the program includes 15 construction problems for a total of 45 tasks, each with easy, medium, and hard settings. Each task requires intermediate steps, such as building scaffolding to hold a block up, for example. This allows testing the AI’s ability to plan its actions.
Note that similar experiments were done with 2 other models, ENHSP And Fast downward. These programs, specializing in sequential operations, failed both tests. Indeed, the first completed almost all the easy and medium problems but none of the hard problems. For the second, he managed only 5 easy problems and 1 medium problem. Steven James Hope that MinePlanner Will be able to solve 45 tasks without human intervention:
We can’t have a human designer come in and tell the AI exactly what it should or shouldn’t do for every task. This is the problem we are trying to address.
To summarize, the program MinePlanner uses Minecraft To test the adaptive capabilities of AI. This is a major problem in the development of autonomous AI, and therefore a project of capital importance. To know more, you can consult the original article in English on the website The new scientist. Don’t hesitate to let us know what you think in the comments on GeekNPlay and all our social networks.