Tor Lattimore

Wednesday 3rd April 2019

Time: 4.00pm

Ground Floor Seminar Room

25 Howland Street, London, W1T 4JG

A generalised information-theoretic approach to sequential decision-making

Balancing exploration and exploitation is a challenging problem in sequential decision-making in the face of uncertainty. Recent developments in the multi-armed bandit literature show that the value of exploration can be quantified in the Bayesian setting using the information gain about the optimal action. In this talk I generalise these ideas and explain how they can be combined with minimax duality and elementary analysis to improve state-of-the-art results for adversarial bandits and related models.

Bio:

Tor Lattimore is research scientist at DeepMind working on multi-armed bandits, reinforcement learning and beyond. Together with Csaba Szepesvari he is the author of the 'Bandit Algorithms' book
(https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fbanditalgs.com&data=02%7C01%7Cb.fong%40ucl.ac.uk%7C4223c4b5422f4b5797b508d6b375db2d%7C1faf88fea9984c5b93c9210a11d9a5c2%7C0%7C0%7C636893716102205728&sdata=XVPdBQegKzdaXYqG0AHCyLNXke52X0AdTQOha8qDyYY%3D&reserved=0). Before joining DeepMind he was an assistant professor at Indiana University.