Collect and Infer

Data-efficiency is a core requirement for artificial intelligence. We are investigating agent architectures and methods, that implement reinforcement learning from scratch based on the principle of 'Collect and Infer' and investigate their application to challenging control and robotic domains. 'Collect and Infer' focuses on the two important aspects of agent learning:

  1. Collecting the 'right data' by advanced exploration methods, like 'Learning by playing' (Riedmiller, ICML 2018)

  2. Effective inference of knowledge from a database of collected transition data, like the 'MPO' algorithm (Maximum A Posteriori Policy Optimisation, Abdolmaleki, 2018) or 'ABM', a highly effective batch RL method (Siegel et. al, ICML 2020)

Martin Riedmiller on The 'Collect and Infer' framework for data-efficient RL

Intro to Collect & Infer

The video is part of a Reinforcement Learning Lecture Series from the University of Alberta, in which I introduce the 'Collect and Infer' perspective on Reinforcement Learning. 'Collect and Infer' is a design principle for data efficient reinforcement learning agents.

Video showing experiments from the 'Learning by Playing' paper, that introduced Scheduled Auxiliary Control (SAC-X) (Riedmiller, ICML 2018)


Learning to play Ball-in-Cup by learning from scratch from raw pixels using Scheduled Auxiliary Control (SAC-X) (Schwab et. al, RSS 2019)

Our 'Learning by Playing' paper explained in the 'Two Minute Papers' series [read the original paper ]


Slides from a keynote talk at EWRL 2018 introducing the 'Collect and Infer' perspective on Reinforcement Learning