Timezone: »

Beyond Target Networks: Improving Deep $Q$-learning with Functional Regularization
Alexandre Piche · Joseph Marino · Gian Maria Marconi · Valentin Thomas · Chris Pal · Mohammad Emtiyaz Khan
Event URL: https://openreview.net/forum?id=JhTPL9BiYy »
A majority of recent successes in deep Reinforcement Learning are based on minimization of square Bellman error. The training is often unstable due to a fast-changing target $Q$-values, and target networks are employed to stabilize by using an additional set of lagging parameters. Despite their advantages, target networks could inhibit the propagation of newly-encountered rewards which may ultimately slow down the training. In this work, we address this issue by augmenting the squared Bellman error with a functional regularizer. Unlike target networks', the regularization here is explicit which not only enables us to use up-to-date parameters but also control the regularization. This leads to a fast yet stable training method. Across a range of Atari environments, we demonstrate empirical improvements over target-network based methods in terms of both sample efficiency and performance. In summary, our approach provides a fast and stable alternative to replace the standard squared Bellman error.

Author Information

Alexandre Piche (Mila)
Joseph Marino (DeepMind)
Gian Maria Marconi (RIKEN)
Valentin Thomas (MILA)
Chris Pal (Montreal Institute for Learning Algorithms, École Polytechnique, Université de Montréal)
Mohammad Emtiyaz Khan (RIKEN)

Emtiyaz Khan (also known as Emti) is a team leader at the RIKEN center for Advanced Intelligence Project (AIP) in Tokyo where he leads the Approximate Bayesian Inference Team. He is also a visiting professor at the Tokyo University of Agriculture and Technology (TUAT). Previously, he was a postdoc and then a scientist at Ecole Polytechnique Fédérale de Lausanne (EPFL), where he also taught two large machine learning courses and received a teaching award. He finished his PhD in machine learning from University of British Columbia in 2012. The main goal of Emti’s research is to understand the principles of learning from data and use them to develop algorithms that can learn like living beings. For the past 10 years, his work has focused on developing Bayesian methods that could lead to such fundamental principles. The approximate Bayesian inference team now continues to use these principles, as well as derive new ones, to solve real-world problems.

More from the Same Authors