Reinforcement Learning with Deep Learning

Lead Research Organisation: University College London
Department Name: Computer Science

Abstract

Advances in RL using deep learning have led to solutions for more complex tasks. There is, however, a need for relatively large amounts of training data. One method to address this issue is to leverage data from different tasks using multi-task learning approaches. The project will take forward this approach to see if this will enable more sample efficient learning and importantly transfer of experience gained on earlier tasks to novel tasks.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509577/1 01/10/2016 24/03/2022
1928590 Studentship EP/N509577/1 01/10/2017 24/09/2021 Tobias Baumann
 
Description The main output so far is my paper on adaptive mechanism design: https://arxiv.org/abs/1806.04067

This paper considers the problem of how an external agent can promote cooperation between artificial learners by distributing additional rewards and punishments based on observing the learners' actions. It proposes a rule for automatically learning how to create the right incentives by considering the players' anticipated parameter updates. Using this learning rule leads to cooperation with high social welfare in matrix games in which the agents would otherwise learn to defect with high probability. The resulting cooperative outcome is stable in certain games even if the planning agent is turned off after a given number of episodes, while other games require ongoing intervention to maintain mutual cooperation. However, even in the latter case, the amount of necessary additional incentives decreases over time.
Exploitation Route Multi-agent reinforcement learning is a cutting-edge research area, and the question of how cooperation can be achieved seems very important if and when artificial learning agents become increasingly widespread - in which case they will interact with both other learning agents and humans in a variety of complex settings.
Sectors Other