Abstract

We have probably all heard the saying “correlation does not imply causation”. But what actually are the differences between a causal and a probabilistic model? How can we learn causal relations from data? And what can causal thinking contribute to machine learning problems such as reinforcement learning?

The first part of this talk (presented by Julius) will be a tutorial-style introduction to causal inference from a statistics and machine learning perspective [1]. After motivating and introducing causal models, we will present different assumptions for causal inference and show how they can all be seen as consequences of a fundamental principle: the principle of independent mechanisms. We will then present different causal discovery methods, with a focus on learning from observational data alone. Time-permitting, we will conclude the first part with connections to semi-supervised and transfer learning.

The second part of the talk (presented by Chaochao) will focus on Causal Reinforcement Learning (Causal RL), which is a promising virgin field and will, without doubt, become an indispensable part of artificial general intelligence. The philosophy behind the integration of Causal Inference and RL is obvious but charming. That is, when looking back at the history of science, human beings always progress in a similar manner to that of Causal RL:

Humans summarize rules or experience from their interaction with nature and then exploit this to improve their adaptation in the next exploration. What CausalRL does is exactly to mimic human behaviors, learning causal relations from an agent communicating with the environment and then optimizing its policy based on the learned causal structures.

In this talk, we will learn what Causal RL is, why we need Causal RL, and how Causal RL works.