What I’m doing now: bots_alive
- I completed my postdoctoral research at the MIT Media Lab in 2014 and am now living in Austin, Texas, building a company that’s developing small robot creatures with the overarching goal of creating a deep illusion of life. Our first product is called bots_alive. Go here to learn more.
- Peter Stone and I published this article in AIJ last year. In it, we dig more deeply into reinforcement learning from human-delivered reward. The article investigates the effect on task performance of changing the temporal discount rate and on changing whether the task is episodic or continuing. We identify issues with non-myopic RL from human reward and then take a step towards addressing those issues, which ultimately may be a critical step in extending these approaches to more complex problems.
- I published an overview of interactive machine learning in AI Magazine. It’s co-authored by Todd Kulesza, Maya Cakmak, and Saleema Amershi, and called Power to the People: The Role of Humans in Interactive Machine Learning. (Author’s copy)
I conduct research on intersections of (1) the control of robots and other agents, (2) machine learning (reinforcement learning in particular), (3) human-computer interaction, and (4) computational models of human behavior for cognitive science research. I’m primarily interested in human interaction with machine learning algorithms, especially when the human fills a teaching role.
In late 2012 I defended my dissertation—Learning from Human-Generated Reward—at UT Austin and joined the MIT Media Lab as a postdoc, working with Cynthia Breazeal’s Personal Robots Group. My research there focused on a project we’re calling Learning from the Wizard, in which a robot learns to emulate its puppeteer’s control, in this case creating an autonomous robot learning companion for young children. At MIT, I also built upon my dissertation research, which was advised by Peter Stone within the LARG research group, and I designed and taught the graduate course Interactive Machine Learning.