Ian Gemp

Ian Gemp
Staff Research Scientist
Google DeepMind
London UK

imgemp at google dot com

I am a Staff Research Scientist working at Google DeepMind (GDM) in London on the Game Theory team. My work focuses on developing equilibrium algorithms (the hammer) and novel game-theoretic formulations of learning problems (the nail). My conference publications have been recognized with awards in areas of eigendecomposition, equilibrium computation, and social choice theory. My research at the intersection of LLMs and game theory has appeared in Wired and Quanta magazine.

Before GDM, I was a PhD student at the College of Information and Computer Sciences at UMass Amherst where I worked with Sridhar Mahadevan as part of the Autonomous Learning Laboratory (ALL) on topics concerning optimization, equilibration, reinforcement learning, multi-agent systems, and dimensionality reduction. My thesis focused on identifying equilibria in differentiable games, specifically those that arise in machine learning applications.

If you're interested in the general evaluation of AI agents. He are two tutorials I gave with Manfred Diaz, Marc Lancot, and Kate Larson on the topic at AAMAS (link) and SAGT (link) 2025.

Selecting the "best" option.
Consider the case where "best" means lowest altitude - find the longitude and latitude of the location with lowest altitude on a contour map.

A point of balance between competing forces.
This encompasses topics such as predator-prey models in biological systems, supply and demand in microeconomics, and even optimization (think about the force of gravity pulling a ball to the point of lowest altitude in a region).

Learning a strategy to maximize a reward.
Imagine teaching a dog a trick or task by feeding it a treat each time it succeeds. Eventually, the dog learns what it needs to do to earn the treat. This area of machine learning draws inspiration from both biology and psychology.

Multiple agents and their environment.
This includes any situation where multiple humans, robots, etc are engaged in a scenario, whether it be competitive, cooperative, or neither so long as their actions affect each other either directly or through the shared environment. It's nearly impossible to think of a situation in real life where this is not the case. As an example, consider two people competing in a game of rock-paper-scissors.

Reducing the number of dimensions.
Dimensionality refers to the number of dimensions, features, attributes, etc (think columns on an excel sheet), so dimensionality reduction just means reducing the number of columns such that you still have enough information on the spreadsheet to complete your task but hopefully much faster now that there's less data to process.

Education

MS, PhD (CS)
Univ of Massachusetts
Amherst MA
2016, 2018

MS (ESAM)
Northwestern Univ
Evanston IL
2011

BS (ME/ESAM)
Northwestern Univ
Evanston IL
2010

High School
Kinkaid
Houston TX
2006

Experience

Google DeepMind Research Scientist
2019 - Present

UMass Amherst
Amherst MA
Research Assistant
2013 - 2018

AWS AI Algorithms
New York NY

Applied Scientist Intern
Summer 2018

Adobe Research
San Jose CA

Data Science Intern
Summer+Fall 2016

Capgemini SA&I
Chicago IL
Consultant
2011-2013

Northwestern Univ
Evanston IL
Research Assistant
2007-2010

UT Medical Center
Houston TX

Research Assistant
Summer 2008/2009