Professor Profile: Sebastian Risi develops AI by having machines play videogames

How do you teach a computer to be creative and adaptive? The same way we humans learn it: by playing games and solving problems. Professor of Digital Design Sebastian Risi uses games to train neural networks for problem solving in the real world. He will give his inaugural lecture entitled "Creative AI: Machines that Play, Adapt, and Create" on August 23.

Sebastian RisiAbout ITUDigital Design Departmentartificial intelligencecomputer games

Written August 4, 2021 8:01 AM by Theis Duelund Jensen

As a student in Germany, when he had to choose an academic field to focus his studies on, Sebastian Risi was torn between two disciplines. Should he go with biology or computer science? On the one hand, he had always been fascinated by computers and the potential of technology, and he had even designed a videogame with a friend in the past. On the other hand, the mysteries of nature, the biological processes that cause cells to develop in perfect patterns, also captured his imagination. In the end, he chose the former at Philipps-Universität Marburg where one of the first classes he took was on neural networks.

“I just found it fascinating and working with neural networks was a way for me to combine my interests in technology and biology. The work I do now is informed by both disciplines,” says the professor, who received his PhD in Computer Science from the University of Central Florida in 2012.

Constructing neural networks is a complex process. The architecture of the network is inspired by how the human brain works, and in its creation, Sebastian Risi and his colleagues look to biological evolution for guidance.

“Nature’s evolution is amazingly creative, just look at the diversity among and variety of organisms. We are trying to come up with algorithms that allow us to create that same variety. Evolution has also allowed really robust organisms to evolve. We try to simulate an environment that will give rise to more creative and adaptive machines. So, we use neural networks, evolutionary algorithms, and deep learning techniques,” says Sebastian Risi, who coordinates work in ITU’s Creative AI Lab and co-directs the Robotics Evolution and Art Lab (REAL) group.

Do neural networks dream of electric sheep?
On Sebastian Risi’s website visitors are met by videos of neural networks constructing habitats in Minecraft, generating new levels in Mario Bros. and recyclable robotic wires in all manner of shapes. They are all examples of his research which is aimed at making artificial intelligence more robust and adaptable by training it to be creative like we humans are. It may all sound like a Philip K. Dick novel about sophisticated artificial intelligence passing as human. Still, we are far from reaching anything near general artificial intelligence, according to the professor.

“You still cannot teach artificial intelligence common sense,” says Sebastian Risi. “Today, we have incredibly complex AI, but it is still restricted to narrowly defined tasks. That is one of the reasons why I am very interested in training neural networks on games, because games are so challenging to machines. They also provide us with a great measure for comparison because we know how humans perform in games.”

In a game setting, an AI is confronted with unforeseen events which is why this line of research is so applicable to real life AI problems: “Take self-driving cars, for instance. There are examples of self-driving cars thinking they are driving towards the horizon, but in reality, it is the back of a truck. This is an example of the network being fooled by unforeseen circumstances. Exploring issues like that in a game context can help make neural networks in general more robust.”

What AI can teach us
On the other side, Sebastian Risi, who in 2019 received a Sapere Aude Starting Grant for his research in adaptive machines in industrial automation, also explores what artificial intelligence can teach us about ourselves. In one project, the professor and his associates used the virtual world of the game Minecraft as a framework in which a neural network formed complex structures – from caterpillars and trees to elaborate interiors. In simplified terms, the artificial intelligence constructs elements by way of locally interacting units – each unit only sees its immediate neighbours – which is exactly how organisms in nature work.

“Our bodies and brains are made by cells communicating with each other, dividing, and growing,” says Sebastian Risi. “We hope simulating this process can help us solve the big biological mysteries and understand how nature can be so precise in its creation.”

But there is also a practical side to this kind of research, he says. If you can train a neural network in Minecraft to think for itself, you may also be able to teach a swarm of a thousand drones to coordinate their movements in unison, or a million nanobots to perform complicated surgery.

“It does sound a bit like science fiction,” Sebastian Risi admits, “but these are the types of things we are trying to lay the groundwork for.”

Professor Sebastian Risi gives his inaugural lecture on Monday, August 23 from 14 – 15:45. At IT University of Copenhagen, Emil Holms Kanal, auditorium 2F13. The lecture is followed by a reception.

Further information