Coevolutionary Deep Reinforcement Learning

Publisher:
IEEE
Publication Type:
Conference Proceeding
Citation:
2020 IEEE Symposium Series on Computational Intelligence, SSCI 2020, 2021, 00, pp. 2600-2607
Issue Date:
2021-01-05
Filename Description Size
Coevolutionary_Deep_Reinforcement_Learning.pdfPublished version218.91 kB
Adobe PDF
Full metadata record
The ability to learn without instruction is a powerful enabler for learning systems. A mechanism for this, selfplay, allows reinforcement learning to develop high performing policies without large datasets or expert knowledge. Despite these benefits, self-play is known to be less sample efficient and suffer unstable learning dynamics. This is in part due to a nonstationary learning problem where an agent's actions influence their opponents and as a consequence the training data they receive. In this paper we demonstrate that competitive pressures can be utilised to improve self-play. This paper leverages coevolution, an evolutionary inspired process in which individuals are compelled to innovate and adapt, to optimise the training of a population of reinforcement learning agents. We demonstrate that our algorithm improves the final performance of a Rainbow DQN trained in the game Connect Four, achieving a 15% higher win percentage over the next leading self-play algorithm. Furthermore, our algorithm exhibits more stable training with less variation in evaluation performance.
Please use this identifier to cite or link to this item: