Causal Decision-Making and Bayesian Optimization

Using causal structure when interventions are expensive and graph knowledge is incomplete.

Summary

This line of work asks how causal assumptions can make sequential optimization and bandit-style exploration more efficient.

Research Question

How should decision-making algorithms use causal graphs, partial graph knowledge, or causal discovery when the goal is to optimize interventions rather than only estimate effects?

Methodology

  • Causal bandits and causal Bayesian optimization
  • Differential Bayesian structure learning
  • Optimization under uncertain or graph-agnostic causal structure

Key Contributions

  • Explores how causal knowledge changes experimental efficiency
  • Connects discovery and decision-making instead of treating them as separate stages
  • Pushes beyond settings that assume the causal graph is fully known in advance

Outputs

  • Publication: (Mukherjee et al., 2024)
  • Related earlier thesis work on scalable causal bandits with adequate causal discovery

References

2024

  1. BDU
    Graph Agnostic Causal Bayesian Optimisation
    Sumantrak Mukherjee, Mengyan Zhang, Seth Flaxman, and 1 more author
    In NeurIPS BDU Workshop, 2024
    Workshop paper, 2024