Causal Bandits
Event details
| Date | 01.11.2022 |
| Hour | 10:30 › 12:30 |
| Speaker | Mikhail Konobeev |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Matthias Grossglauser
Thesis advisor: Prof. Negar Kiyavash
Co-examiner: Prof. Michael Kapralov
Abstract
We consider the problem of maximizing a designated reward variable in probabilistic causal model via interventions on other variables. Our goal is to eventually develop methods that do not require the knowledge of the underlying causal graph. We present three papers that are the benchmarks for our future work: one that defines the causal bandit problem, another that solves it using only the knowledge of the essential graph, and the last that characterizes the set of actions an agent has to explore. We present a randomized algorithm for finding the parent of the reward node and distinguish three classes of graphs with the asymptotic number of interventions of the proposed algorithm ranging from logarithmic to sublinear to linear. We also discuss different other settings in which we would like to achieve our goal.
Background papers
1. Causal bandits: Learning good interventions via causal inference
2. Causal bandits with unknown graph structure
3. Structural causal bandits: where to intervene?
Exam president: Prof. Matthias Grossglauser
Thesis advisor: Prof. Negar Kiyavash
Co-examiner: Prof. Michael Kapralov
Abstract
We consider the problem of maximizing a designated reward variable in probabilistic causal model via interventions on other variables. Our goal is to eventually develop methods that do not require the knowledge of the underlying causal graph. We present three papers that are the benchmarks for our future work: one that defines the causal bandit problem, another that solves it using only the knowledge of the essential graph, and the last that characterizes the set of actions an agent has to explore. We present a randomized algorithm for finding the parent of the reward node and distinguish three classes of graphs with the asymptotic number of interventions of the proposed algorithm ranging from logarithmic to sublinear to linear. We also discuss different other settings in which we would like to achieve our goal.
Background papers
1. Causal bandits: Learning good interventions via causal inference
2. Causal bandits with unknown graph structure
3. Structural causal bandits: where to intervene?
Practical information
- General public
- Free