Improving Stability in Deep Reinforcement Learning with Weight Averaging

By Evgenii Nikishin, Pavel Izmailov, Ben Athiwaratkun, Dmitrii Podoprikhin, Timur Garipov, Pavel Shvechikov, Dmitry P. Vetrov, Andrew Gordon Wilson. 2018

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Pavel Izmailov
2. Andrew Gordon Wilson
3. Wesley J. Maddox
4. Jean Kaddour
5. J'er'emy Scheurer
6. Dmitry P. Vetrov
7. Timur Garipov
8. Anca Dragan

Citations in the corpus

• A Simple Baseline for Bayesian Uncertainty in Deep Learning (2019, background): “SWA and related ideas have been successfully applied to a wide range of applications [see e.g. 2, 61, 62, 51].”
• Questions for Flat-Minima Optimization of Modern Neural Networks (2022, background): “We select these due to their scalability to large models and excellent results across a variety of learning tasks (Nikishin et al., 2018; Athiwaratkun et al., 2019; Chen et al., 2021a; Bahri et al., 2021).”
• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback (2023, background): “Deep RL is unstable, and results are often highly sensitive to initialization and difficult to reproduce [Nikishin et al., 2018, Irpan, 2018, Henderson et al., 2018].”