Suggestions
Alex Olshevsky
Associate Professor at Boston University working on reinforcement learning and optimization. Founder of a startup specializing in recommendation systems.
Alex Olshevsky is an Associate Professor at Boston University with multiple affiliations within the university.123 He holds positions in the Department of Electrical and Computer Engineering, the Division of Systems Engineering, and the Department of Computer Science.36
Academic Focus
Olshevsky's research interests include:
- Distributed optimization, particularly in machine learning
- Control of networks and multi-agent systems
- Reinforcement learning
- Control theory
- Optimization46
Professional Details
- Office Location: 8 St. Mary's St Boston, MA 02215, PHO 419
- Email: [email protected]
- Phone: (617) 353-01732
Additional Roles
Beyond his primary academic positions, Olshevsky is also:
- An Affiliate Professor of Systems Engineering at Boston University5
- Associated with the Rafik Hariri Institute for Computing and Computational Science & Engineering at BU3
Entrepreneurial Ventures
In addition to his academic work, Alex Olshevsky is the founder of a startup that specializes in recommendation systems.7
Alex Olshevsky's diverse roles and research interests demonstrate his expertise in various aspects of computer science, engineering, and systems optimization, making him a valuable asset to Boston University's academic community.
Highlights
I am co-organizing a "Reinforcement Learning Day" at BU on May 10:
https://lnkd.in/ejcZs9-5
We have a great lineup of speakers who will talk about recent progress in RL, touching on both theory as well as applications of RL. Registration and attendance is free. Please re-share and consider attending!
Very excited to share this paper https://lnkd.in/diC6zm_b with Haoxing Tiang and Yannis Paschalidis which is scheduled appear in ICLR 2023 in Kigali. Below, I'll give a quick two-paragraph summary of the result.
While merging neural networks and reinforcement learning has achieved spectacular results over the last decade, relatively few theoretical guarantees are available, and in practice the most popular algorithms often diverge. Our high-level goal is to develop methods which rely on neural networks but simultaneously retain the nice theoretical guarantees associated with more classical RL methods.
Our main result is quite simple: if we use a neural network of width m for policy evaluation -- training it with projected temporal difference learning, where the projection is onto a ball B of constant radius around the initial condition -- then this approximates the true value function to an error of O(epsilon + 1/m^{1/2}) where epsilon is an approximation error associated with the best neural network in B. It is natural that the error goes to zero with network width m, since by NTK theory neural networks get more linear as they get wide; but getting concrete bounds out of this for finite m is much more challenging. In particular, the network could behave quite nonlinearly during training for moderately large m and our approximation results would still apply.
As always, very happy to receive feedback and comments on this work. #learning #neuralnetworks
