Combating False Negatives in Adversarial Imitation Learning (Student Abstract)

  • Konrad Żołna Jagiellonian University
  • Chitwan Saharia IIT Bombay
  • Leonard Boussioux Mila
  • David Yu-Tung Hui Mila
  • Maxime Chevalier-Boisvert Mila
  • Dzmitry Bahdanau Mila
  • Yoshua Bengio Mila

Abstract

We define the False Negatives problem and show that it is a significant limitation in adversarial imitation learning. We propose a method that solves the problem by leveraging the nature of goal-conditioned tasks. The method, dubbed Fake Conditioning, is tested on instruction following tasks in BabyAI environments, where it improves sample efficiency over the baselines by at least an order of magnitude.

Published
2020-04-03
Section
Student Abstract Track