Timezone: »
We propose a simple architecture for deep reinforcement learning by embedding inputs into a learned Fourier basis and show that it improves the sample efficiency of both state-based and image-based RL. We perform infinite-width analysis of our architecture using the Neural Tangent Kernel and theoretically show that tuning the initial variance of the Fourier basis is equivalent to functional regularization of the learned deep network. That is, these learned Fourier features allow for adjusting the degree to which networks underfit or overfit different frequencies in the training data, and hence provide a controlled mechanism to improve the stability and performance of RL optimization. Empirically, this allows us to prioritize learning low-frequency functions and speed up learning by reducing networks' susceptibility to noise in the optimization process, such as during Bellman updates. Experiments on standard state-based and image-based RL benchmarks show clear benefits of our architecture over the baselines.
Author Information
Alex Li (Carnegie Mellon University)
Deepak Pathak (Carnegie Mellon University)
More from the Same Authors
-
2021 : RB2: Robotic Manipulation Benchmarking with a Twist »
Sudeep Dasari · Jianren Wang · Joyce Hong · Shikhar Bahl · Yixin Lin · Austin Wang · Abitha Thankaraj · Karanbir Chahal · Berk Calli · Saurabh Gupta · David Held · Lerrel Pinto · Deepak Pathak · Vikash Kumar · Abhinav Gupta -
2021 : The CLEAR Benchmark: Continual LEArning on Real-World Imagery »
Zhiqiu Lin · Jia Shi · Deepak Pathak · Deva Ramanan -
2021 : Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives »
Murtaza Dalal · Deepak Pathak · Russ Salakhutdinov -
2022 : Test-time adaptation with slot-centric models »
Mihir Prabhudesai · Sujoy Paul · Sjoerd van Steenkiste · Mehdi S. M. Sajjadi · Anirudh Goyal · Deepak Pathak · Katerina Fragkiadaki · Gaurav Aggarwal · Thomas Kipf -
2022 : Test-time adaptation with slot-centric models »
Mihir Prabhudesai · Sujoy Paul · Sjoerd van Steenkiste · Mehdi S. M. Sajjadi · Anirudh Goyal · Deepak Pathak · Katerina Fragkiadaki · Gaurav Aggarwal · Thomas Kipf -
2022 Poster: Continual Learning with Evolving Class Ontologies »
Zhiqiu Lin · Deepak Pathak · Yu-Xiong Wang · Deva Ramanan · Shu Kong -
2021 Oral: Interesting Object, Curious Agent: Learning Task-Agnostic Exploration »
Simone Parisi · Victoria Dean · Deepak Pathak · Abhinav Gupta -
2021 Poster: Accelerating Robotic Reinforcement Learning via Parameterized Action Primitives »
Murtaza Dalal · Deepak Pathak · Russ Salakhutdinov -
2021 Poster: Discovering and Achieving Goals via World Models »
Russell Mendonca · Oleh Rybkin · Kostas Daniilidis · Danijar Hafner · Deepak Pathak -
2021 Poster: Interesting Object, Curious Agent: Learning Task-Agnostic Exploration »
Simone Parisi · Victoria Dean · Deepak Pathak · Abhinav Gupta -
2020 Poster: Neural Dynamic Policies for End-to-End Sensorimotor Learning »
Shikhar Bahl · Mustafa Mukadam · Abhinav Gupta · Deepak Pathak -
2020 Spotlight: Neural Dynamic Policies for End-to-End Sensorimotor Learning »
Shikhar Bahl · Mustafa Mukadam · Abhinav Gupta · Deepak Pathak -
2020 Session: Orals & Spotlights Track 14: Reinforcement Learning »
Deepak Pathak · Martha White -
2020 Poster: Sparse Graphical Memory for Robust Planning »
Scott Emmons · Ajay Jain · Misha Laskin · Thanard Kurutach · Pieter Abbeel · Deepak Pathak