Artificial Intelligence (AI) may have beaten professional Go player Lee Sedol, but it isn’t done playing games with us yet.
Now, the software engineers at Microsoft Research are taking on an even more ambitious project. Where Google’s DeepMind took on the board game world, Microsoft’s AI plans to take on video games.
The team’s Project AIX will reportedly be capable of teaching itself to play Minecraft.
Sandboxing Artificial Intelligence
Minecraft is an open-world creation game for consoles and PCs where players can dig, build and take down creepers to their heart’s content.
AIX will need to learn to climb and swim, as well as to make and use tools, in order to traverse a wide variety of terrain.
The game also has no real set tasks or goals, so by placing the AI within this world of endless possibilities, the team hopes to teach it some common sense and decision-making skills.
“We’re trying to program it to learn, as opposed to programming it to accomplish certain tasks,” said Fernando Diaz, a senior researcher at DeepMind.
A Machine Learning a Game
The AI agent will be placed into a Minecraft game with no concept of its surroundings or purpose. Much like a toddler, it will be expected to learn how to move and interact with its environment via trial and error. It will also need to learn to understand when it has achieved its current goal.
It will need to identify which aspects of gameplay are immediately relevant (such as time of day and nearby enemies) and set priorities accordingly.
As a bonus, placing the AI into a gameplay setting like this will allow the team to test how the AI handles being contained within an avatar.
“It allows you to have ‘embodied’ AI,” said Matthew Johnson, AIX software engineer. “Rather than have a situation where the AI sees an avatar of itself, it can actually be inside—looking out through the eyes of something that is living in the [Minecraft] world. We think this is an essential part of building this kind of general intelligence.”
Microsoft hasn't announced any specific test requirements for AIX, but given that Minecraft is a building game, the AI will need to have some capacity for creativity.
One question that remains with this project is the level of creativity that the AI will need to have. If the team wants to have the AI build something, it will need to have a vision of its intended creation and a plan for building it—including decisions like material selection and block placement.
Applications for Autodidactic Artificial Intelligence
Deep learning systems are popular with platforms like Facebook and Google because they provide insights with commercial applications. While this type of AI is capable of activities like identifying and labeling images on its own, it is still lacking in what humans consider common sense.
Much like other attempts to recreate neural networks with computers, the AIX project has the potential to improve a computer’s understanding of what AI researchers call general intelligence. This is comparable to the complex methods of learning and decision-making used by humans.
The team at DeepMind is hoping that a tool like AIX could unlock new ways for AI to learn about its environment. It could also serve as a stepping stone to a human-AI interface for use in industrial robots.
The team is currently using AIX for its own research purposes, but the project will be available via an open-source license as early as summer 2016.
For more information about Project AIX, check out the Microsoft AI webpage.