- author = "John D. Co-Reyes and Yingjie Miao and Daiyi Peng and Esteban Real and Sergey Levine and Quoc V. Le and Honglak Lee and Aleksandra Faust",
- title = "Evolving Reinforcement Learning Algorithms",
- howpublished = "ArXiv",
- year = "2021",
- month = "8 " # jan,
- keywords = "genetic algorithms, genetic programming, genetic improvement, computer video games",
- eprint = "2101.03958",
- archiveprefix = "arXiv",
- primaryclass = "cs.LG",
-
URL = "
https://arxiv.org/abs/2101.03958",
-
video_url = "https://slideslive.com/38941399/evolving-reinforcement-learning-algorithms",
- size = "15 pages",
- abstract = "We propose a method for meta-learning reinforcement learning algorithms by searching over the space of computational graphs which compute the loss function for a value-based model-free RL agent to optimize. The learned algorithms are domain-agnostic and can generalize to new environments not seen during training. Our method can both learn from scratch and bootstrap off known existing algorithms, like DQN, enabling interpretable modifications which improve performance. Learning from scratch on simple classical control and gridworld tasks, our method rediscovers the temporal-difference (TD) algorithm. Bootstrapped from DQN, we highlight two learned algorithms which obtain good generalization performance over other classical control tasks, gridworld type tasks, and Atari games. The analysis of the learned algorithm behaviour shows resemblance to recently proposed RL algorithms that address overestimation in value-based methods.",
-
notes = "Research at Google, Mountain View, CA 94043,
USA
'bootstrapping off existing algorithms'. 'The search is done over 300 CPUs and run for roughly 72 hours, at which point around 20000 programs have been evaluated.'",
