Reproduce “Privileged Information Dropout in Reinforcement Learning” on OpenResearch

About this paper

Pierre-Alexandre Kamienny, Kai Arulkumaran, Feryal Behbahani, Wendelin Boehmer, Shimon Whiteson

A framework called Privileged Information Dropout (PI-Dropout) is adapted for reinforcement learning, allowing agents to utilize extra information during training that is not required for deployment. In a partially observable gridworld, this approach consistently outperformed distillation and auxiliary tasks, enabling more robust navigation.

Key insights

• PI-Dropout provides a principled mechanism to integrate privileged information directly into the input features of both policy and value functions during RL training.
• Conditioning the multiplicative dropout noise on the privileged information is crucial, leading to superior performance compared to naive dropout or standard information dropout.
• The method can effectively utilize diverse forms of privileged information, such as full state or sub-goals, demonstrating its versatility in different task settings.

Summary by alphaXiv.