Dan Bethell

My PhD focuses on safe artificial intelligence, specifically the safe exploration of reinforcement learning agents. This allows the agent to learn and explore a task or environment without risk or harm, like guiding a child to play safely. It ensures the agent will make smart decisions within the limits of safety. This is especially crucial in high-stakes applications like robot navigation, where errors can have severe consequences.

There are a number of factors that influenced me to study for my PhD here at York. The most salient is the opportunity to be supervised by leading experts in safety and autonomous systems. Working alongside them has given me the backing I need to become a great researcher.

The University itself also shares the same ethos towards research for the public good, as seen in their recent advances in furthering research towards safe autonomy through their grants and opening of the Institute for Safe Autonomy.

Within my PhD, I was lucky enough to visit Vancouver, Canada after my work was accepted into an A* conference, AAAI’24. This gave me the chance to present my research and meet with world-leading academics in my area.

What I love about my work is that reinforcement learning aims to find a mathematically optimal solution to a given task. It becomes fascinating when it finds a solution that seemed non-intuitive from our human perspective. A significant example of this is ‘Move 37’ in a match between AlphaGo, a deep reinforcement learning agent trained to play Go, against Lee Sedol, a world-leading Go player. This move which seemed unconventional to a human player, turned out to be a game-deciding strategy. It showed how powerful reinforcement learning agents can be and how they can deviate from human understanding to achieve
superhuman performance.