PhD Thesis Proposal: Kate Avery, Safe Decision-Making under Causal Uncertainty 

Content

Speaker: 

Kate Avery

Abstract:

This proposal examines safe decision-making under causal uncertainty. Causal uncertainty is defined as uncertainty from unknown causal structures (what variables cause what other variables) and mechanisms (how variables cause other variables). Safe decision-making is the ability to make decisions and learn policies that maintain formal guarantees on system behavior under reasonable assumptions. We explore safe decision-making in both offline and online settings. That is, we explore how to make safe decisions using only previously collected data, as well as how to make safe decisions when we can perform targeted interventions to find out more about our environment.

The first part of the proposal investigates safe decision-making in an offline setting. We create a method for evaluating and learning robust decision policies under uncertain causal mechanisms, specifically the worst-case realistic mechanisms. The latter half of the proposal examines safe decision-making in the online setting. We first propose a method for learning and representing causal models using Bayesian optimization and probabilistic programming. Safety constraints are then incorporated through safe Bayesian optimization, which restricts exploration to a set of safe states. Finally, the proposal extends safe Bayesian optimization to counterfactual policy constraints. These constraints require that the interventions taken by Bayesian optimization are safe relative to a baseline policy, and they are very common when learning causal structures and mechanisms from experimental and observational data.

Advisor:

David Jensen