Artificial Intelligence

Feature Story

Want to Teach An AI Novelty? First, Teach It Monopoly. Then Throw Out the Rules

July 31, 2020

Alex Baker

Researchers from the USC Viterbi School of Engineering's Information Sciences Institute (ISI) have partnered with Purdue University to take part in the Defense Advanced Research Projects Agency (DARPA)-funded program that seeks to develop the science that will allow AI systems to adapt to novelty, or new conditions that haven't been seen before.

Take an AI that has been trained to play a standard game of Monopoly. What if you change the rules so that you can buy houses and hotels without first getting a monopoly? What if the game is set to end after 100 turns instead of waiting for bankruptcies? These are both novelties which would affect the optimal strategy to win.

And yet, as Mayank Kejriwal, the primary investigator on the project and a USC Viterbi research assistant professor, added, even today the most advanced AIs are ill-equipped to deal with this sort of novelty.

"Even though there have been lots of advancements in AI, they are very task specific," Kejriwal said. "The moment you introduce changes that the AI is not specifically equipped to handle, you have to go back and retrain the program. There is no general AI, something that can adapt to novel situations. We are really in uncharted waters because there is no science of novelty."

"That's the significance of this project," he added. "It's not just about improving some specific AI module. By developing a science of novelty, we are laying the foundation for future generations of AI."

The Science of Artificial Intelligence and Learning for Open-world Novelty (SAIL-ON) program, or SAIL-ON program began in November of 2019 and will continue until 2023. At the program's end, the Department of Defense hopes to use the research in a range of applications, from autonomous disaster-relief robots to self-driving military vehicles. The USC and Purdue collaborative team has been allocated $1.2 million  from DARPA, and will likely receive more as the program goes on.

In some respects, AI has already surpassed human capabilities. Kejriwal cited AlphaZero as an example - a computer program that uses machine learning to play board games such as chess and Go, can now beat even the most advanced human players.

Read More

Groups:

News

USC Viterbi Sustainability Projects Receive $150,000 Funding Boost

August 18, 2020

USC Viterbi researchers will lead three new projects to create innovative solutions to our most urgent environmental problems, thanks to $150,000 in funding from the USC Center for Sustainability Solutions.

Read More

Want to Teach An AI Novelty? First, Teach It Monopoly. Then Throw Out the Rules.

July 31, 2020

Researchers from the USC Viterbi School of Engineering's Information Sciences Institute (ISI) have partnered with Purdue University to take part in the Defense Advanced Research Projects Agency (DARPA)-funded program that seeks to develop the science that will allow AI systems to adapt to novelty, or new conditions that haven't been seen before.

Read More

USC ISI Researchers Develop Improv-Based Chatbots

July 1, 2020

What would conversations with Alexa be like if she was a regular at The Second City?

Read More
See More Stories »

Feature Story

Want to Teach An AI Novelty? First, Teach It Monopoly. Then Throw Out the Rules

July 31, 2020

Alex Baker

Researchers from the USC Viterbi School of Engineering's Information Sciences Institute (ISI) have partnered with Purdue University to take part in the Defense Advanced Research Projects Agency (DARPA)-funded program that seeks to develop the science that will allow AI systems to adapt to novelty, or new conditions that haven't been seen before.

Take an AI that has been trained to play a standard game of Monopoly. What if you change the rules so that you can buy houses and hotels without first getting a monopoly? What if the game is set to end after 100 turns instead of waiting for bankruptcies? These are both novelties which would affect the optimal strategy to win.

And yet, as Mayank Kejriwal, the primary investigator on the project and a USC Viterbi research assistant professor, added, even today the most advanced AIs are ill-equipped to deal with this sort of novelty.

"Even though there have been lots of advancements in AI, they are very task specific," Kejriwal said. "The moment you introduce changes that the AI is not specifically equipped to handle, you have to go back and retrain the program. There is no general AI, something that can adapt to novel situations. We are really in uncharted waters because there is no science of novelty."

"That's the significance of this project," he added. "It's not just about improving some specific AI module. By developing a science of novelty, we are laying the foundation for future generations of AI."

The Science of Artificial Intelligence and Learning for Open-world Novelty (SAIL-ON) program, or SAIL-ON program began in November of 2019 and will continue until 2023. At the program's end, the Department of Defense hopes to use the research in a range of applications, from autonomous disaster-relief robots to self-driving military vehicles. The USC and Purdue collaborative team has been allocated $1.2 million  from DARPA, and will likely receive more as the program goes on.

In some respects, AI has already surpassed human capabilities. Kejriwal cited AlphaZero as an example - a computer program that uses machine learning to play board games such as chess and Go, can now beat even the most advanced human players.

Read More

Groups:

Events

Unless otherwise noted, seminars are open to the public.

Oct 30 Judy Hoffman, Georgia TechAI Seminar

Understanding and Mitigating Bias in Visual Recognition

11:00am - 12:00pm PDT
Nov 05 Elaheh Raisi (Brown University)AI Seminar

Learning with less dependency on labeled data for cyberbullying detection and image classification tasks

10:00am - 11:00am PDT
Nov 12 John Hewitt (Stanford Univ)NL Seminar

The Unreasonable Syntactic Expressivity of RNNs

11:00am - 12:00pm PDT
See More Events »