Neural Architecture Search under black-box objectives with deep reinforcement learning and increasingly-sparse rewards

Mitchel Alioscha-Perez, Abel Diaz Berenguer, Ercheng Pei, Meshia Cédric Oveneke, Hichem Sahli

Research output: Chapter in Book/Report/Conference proceedingConference paper

1 Citation (Scopus)

Abstract

In this paper, we address the problem of neural architecture search (NAS) in a context where the optimality policy derivatives. In this scenario, O(A) typically provides readings from a set of sensors on how a neural network architecture A fares in a target hardware, including its: power consumption, working temperature, cpu/gpu usage, central bus occupancy, and more. Current differentiable NAS approaches fail in this problem context due to lack of access to derivatives, whereas traditional reinforcement learning NAS approaches remain too expensive computationally. As solution, we propose a reinforcement learning NAS strategy based on policy gradient with increasingly sparse rewards. We rely on the fact that one does not need to fully train the weights of two neural networks to compare them. Our solution starts by comparing architecture candidates with almost fixed weights and no training, and progressively shifts toward comparisons under full weights training. Experimental results confirmed both the accuracy and training efficiency of our solution, as well as its compliance with soft/hard constraints imposed on the sensors feedback. Our strategy allows finding near-optimal architectures significantly faster, in approximately 1/3 of the time it would take otherwise.
Original languageEnglish
Title of host publicationInternational Conference on Artificial Intelligence in Information and Communication (ICAIIC)
PublisherIEEE
Pages276-281
Number of pages6
ISBN (Electronic)978-1-7281-4985-1
ISBN (Print)978-1-7281-4986-8
DOIs
Publication statusPublished - 21 Feb 2020
EventInternational Conference on Artificial Intelligence in Information and Communication: ICAIIC2020 - Takakura Hotel, Fukuoka, Japan
Duration: 19 Feb 202021 Feb 2020
http://icaiic.org/

Conference

ConferenceInternational Conference on Artificial Intelligence in Information and Communication
Abbreviated titleICAIIC2020
CountryJapan
CityFukuoka
Period19/02/2021/02/20
Internet address

Fingerprint

Dive into the research topics of 'Neural Architecture Search under black-box objectives with deep reinforcement learning and increasingly-sparse rewards'. Together they form a unique fingerprint.

Cite this