David Abel

Welcome!

I am a Research Scientist at DeepMind in London. Before that, I completed my Ph.D in Computer Science and Masters in Philosophy at Brown University, where I was fortunate to be advised by Prof. Michael Littman (CS), and Prof. Joshua Schechter (Philosophy).

News

• Dec. 2021: Held a mentoring session via mementor.
• Nov. 2021: Honored that our NeurIPS paper was awarded "Outstanding Paper"!
• Oct. 2021: Our paper "On the Expressivity of Markov Reward" was accepted to NeurIPS.
• Oct. 2021: Talk at Mila.
• Oct. 2021: Guest lecture in Chris Tanner's capstone course at Harvard.
• Sep. 2021: I am getting married! I will be away from email September 15 - October 6, 2021.
• July 2021: Held a mentoring session via mementor.
• July 2021: Spotlight at the RL Theory workshop at ICML on our paper, Bad-Policy Density.
• July 2021: Talk at Purdue.
• May 2021: New preprint out on Control of Representations, led by Mark Ho.
• May 2021: Co-hosting a panel on the philosophy of AGI at ICLR with Jacob Beck and Clare Lyle.
• Apr. 2021: New ICML paper out on Peng’s Q(λ), led by Tadashi Kozuno and Yunhao Tang.
• Feb. 2021: Guest lecture in Chris Tanner's capstone course at Harvard.
• Dec. 2020: Talk(s) in GRLA.
• Dec. 2020: Lipschitz Lifelong RL accepted to AAAI.
• Oct. 2020: I joined DeepMind London as a Research Scientist.
• Sep. 2020: I gave a guest lecture in Chris Tanner's capstone course at Harvard.
• Sep. 2020: I moved to London.
• Aug. 2020: New blog post with collaborators summarizing our AISTATS paper.
• June 2020: New ICML paper out on affordances in RL.
• Apr. 2020: My dissertation is done!
• Apr. 2020: I defended my thesis.
• Mar. 2020: New AISTATS paper out on state-action abstraction.

Interests

My research focuses on bringing clarity to the central philosophical questions surrounding computation and learning.

I value research that concentrates on providing new understanding, and tend to get excited by simple but foundational questions. I typically work with the reinforcement learning problem, drawing on tools and perspectives from computational learning theory, computational complexity, and analytic philosophy.

I am currently interested in better defining the AI problem. Previously, my dissertation studied how effective agents model the worlds they inhabit, focusing on the representational practices that underly effective learning and planning.

Featured Research

	On the Expressivity of Markov Reward NeurIPS 2021 We study the expressivity of Markov reward functions in finite environments by analysing what kinds of tasks such functions can express. Joint work with Will Dabney, Anna Harutyunyan, Mark K. Ho, Michael L. Littman, Doina Precup, Satinder Singh.
	A Theory of Abstraction in Reinforcement Learning Ph.D Thesis, 2020 My dissertation, aimed at understanding abstraction and its role in effective reinforcement learning. Advised by Michael L. Littman.
	Value Preserving State-Action Abstractions AISTATS 2020 We prove which combinations of state abstractions and options are guaranteed to preserve representation of near-optimal policies in any finite Markov Decision Process. Joint work with Nathan Umbanhowar, Khimya Khetarpal, Dilip Arumugam, Doina Precup, and Michael L. Littman.
	Lipschitz Lifelong Reinforcement Learning AAAI 2021 We examine the Lipschitz continuity of value functions and MDPs, then exploit these properties to develop a PAC-MDP algorithm for lifelong RL called Lipschitz RMax. Led by Erwan Lecarpentier, joint with Kavosh Asadi, Yuu Jinnai, Emmanuel Rachelson, and Michael L. Littman.
	What can I do here? A Theory of Affordances in Reinforcement Learning ICML 2020 We develop a theory of affordances in the context of RL and planning. Led by Khimya Khetarpal, joint with Zafarali Ahmed, Gheorghe Comanici, and Doina Precup.
	The Efficiency of Human Cognition Reflects Planned Use of Information Processing AAAI 2020 We develop a model that characterizes the planned use of information processing as a meta-reasoning problem and study this model's capacity to predict human reaction times in simple tasks. By Mark K. Ho, Jonathan D. Cohen, Michael L. Littman, Thomas L. Griffiths.
	The Value of Abstraction Current Opinions in Behavioral Science 2019 We discuss the vital role that abstraction plays in efficient decision making. Led by Mark K. Ho, joint with Michael L. Littman, Thomas L. Griffiths.
	The Expected-Length Model of Options IJCAI 2019 We introduce and motivate the Expected-Length Model of Options, a simpler alternative for characterizing the transition and reward functions of options. Joint with John Winder, Marie desJardins, and Michael L. Littman.
	State Abstraction as Compression in Apprenticeship Learning AAAI 2019 We study state abstractions that trade-off between compression and optimality through rate-distortion theory. Joint work with Dilip Arumugam, Kavosh Asadi, Yuu Jinnai, Michael L. Littman, and Lawson L.S. Wong.
	Finding Options that Minimize Planning Time ICML 2019 We prove that the problem of finding options that minimize planning time is NP-Hard. Led by Yuu Jinnai, joint with D Ellis Hershkowitz, Michael L. Littman, and George Konidaris.

Personal

For fun, I'm a big fan of basketball, snowboarding, games, and music (I play guitar/piano/violin and mostly listen to progressive metal). I now live in London, UK, with my wife Elizabeth and our dog Barley.

Always up for a chat -- shoot me an email if you'd like to discuss anything!